Automatic Bug Bounty

Masaryk University
Faculty of Informatics
Automating Bug Bounty
Master’s Thesis
Ján Masarik
Brno, Spring 2019

Replace this page with a copy of the official signed thesis assignment and the
copy of the Statement of an Author.
Declaration
Hereby I declare that this paper is my original authorial work, which
I have worked out on my own. All sources, references, and literature
used or excerpted during elaboration of this work are properly cited
and listed in complete reference to the due source.
Ján Masarik
Advisor: prof. RNDr. Václav Matyáš M.Sc. Ph.D.
i
Acknowledgement
First, I would like to express huge gratitude to my advisor prof. RNDr.
Václav Matyáš M.Sc. Ph.D. for the continuous guidance, personal
approach, open-mindedness and for his interesting lectures that were
one of the reasons I decided to study cybersecurity.
My thanks also goes to Anshuman Bhartiya, creator of the Bounty-
Machine framework, for the idea and consultations that guided me
through the initial steps of the implementation.
I would also like to thank my colleagues from Kiwi.com, especially
to Stanislav Komanec for the encouragement and to Alex Viscreanu
for his support and knowledge.
Last but not least, I would like to thank my family and girlfriend
for their support and patience.
ii
Abstract
This thesis introduces the reasoning behind the increasing popularity
of the bug bounty programs and analyzes the most common, impact-
ful bugs whose discovery can be automated at scale. As there was no
suitable existing solution for similar automation identified, the thesis
introduces a new framework, specifically designed for the bug bounty
automation use case. Particular attention of the thesis is dedicated
the usable security of the object storage console on six major cloud
providers, and a large scale quantitative research that revealed dozens
of misconfigured Azure Blob Storage buckets containing highly sen-
sitive data. A set of recommendations, based on the results of the
research, was formulated and delivered to each of the tested cloud
providers.
iii
Keywords
automation, bug bounty, Kubernetes, penetration testing, security
iv
Contents
1 Introduction 1
2 Bug bounty 3
2.1 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Benefits for the organization . . . . . . . . . . . . . . . . . 4
2.3 Comparison with traditional approaches . . . . . . . . . . . 4
2.4 Self-hosted bug bounty program . . . . . . . . . . . . . . . 6
2.5 Bug bounty platforms . . . . . . . . . . . . . . . . . . . . . 7
2.5.1 HackerOne . . . . . . . . . . . . . . . . . . . . . 7
2.5.2 Bugcrowd . . . . . . . . . . . . . . . . . . . . . . 9
2.5.3 Others . . . . . . . . . . . . . . . . . . . . . . . . 9
3 Common bugs in bug bounty 11

3.1 Object storage misconfiguration . . . . . . . . . . . . . . . 11
3.1.1 Amazon Web Services S3 . . . . . . . . . . . . . 13
3.1.2 DigitalOcean Spaces . . . . . . . . . . . . . . . . 15
3.1.3 Azure Blob Storage . . . . . . . . . . . . . . . . . 17
3.1.4 Google Cloud Storage . . . . . . . . . . . . . . . 18
3.1.5 Alibaba Object Storage Service . . . . . . . . . . 20
3.1.6 Oracle Cloud Object Storage . . . . . . . . . . . . 21
3.2 Leaked sensitive API keys . . . . . . . . . . . . . . . . . . . 23
3.3 Broken Authentication . . . . . . . . . . . . . . . . . . . . 24
3.4 Security Misconfiguration . . . . . . . . . . . . . . . . . . 25
3.5 Subdomain takeover . . . . . . . . . . . . . . . . . . . . . . 26
3.6 Server Side Request Forgery . . . . . . . . . . . . . . . . . 27
4 Bug bounty automation 29

4.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Existing solutions . . . . . . . . . . . . . . . . . . . . . . . 30
4.3 Technology used . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3.1 Docker . . . . . . . . . . . . . . . . . . . . . . . . 32
4.3.2 Kubernetes . . . . . . . . . . . . . . . . . . . . . 32
4.3.3 Argoproj . . . . . . . . . . . . . . . . . . . . . . . 33
4.3.4 GitLab . . . . . . . . . . . . . . . . . . . . . . . . 36
5 Proposed solution 37
v
5.1 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.2 Triggers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.2.1 Manual . . . . . . . . . . . . . . . . . . . . . . . . 39
5.2.2 Bugtab . . . . . . . . . . . . . . . . . . . . . . . . 40
5.3 Workflows . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.3.1 Subdomain enumeration . . . . . . . . . . . . . . 43
5.3.2 Bucket enumeration . . . . . . . . . . . . . . . . 44
5.3.3 Subdomain takeover . . . . . . . . . . . . . . . . 46
5.3.4 Git secrets . . . . . . . . . . . . . . . . . . . . . . 46
5.4 Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.5 Alerting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.6 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6 Misconfigured Azure Blob Storage 50

6.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.1.1 Forward DNS . . . . . . . . . . . . . . . . . . . . 50
6.1.2 Alexa top 1 million . . . . . . . . . . . . . . . . . 51
6.2 Enumeration workflow . . . . . . . . . . . . . . . . . . . . 51
6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.4 Recommendations . . . . . . . . . . . . . . . . . . . . . . . 53
7 Conclusions 55
Appendices 69
A Archive structure 70
B Vulnerability disclosure email 72
C DigitalOcean Spaces enumeration workflow 73
vi
1 Introduction
The bug bounty industry has experienced a significant growth in the
past few years. Now, more and more companies are running their
bug bounty program and accepting reports from hackers1 around
the world [1]. In Chapter 2, the bug bounty phenomenon will be
introduced and discussed from the perspective of a hacker and a
company running a program, along with a comparison with traditional
application security approaches such as penetration testing.
Widespread, high impact bugs in bug bounty, along with the op-
tions of their automated discovery are analyzed in Chapter 3. A signif-
icant part of this chapter is focused on object storage misconfiguration
bugs, that were responsible for a substantial amount of data breaches
that affected companies such as Verizon, Viacom or Accenture [2, 4, 3],
along with U.S. agencies NSA and NGA2 [5, 6] in the past two years.
The usable security of object storage console was analyzed on six
different cloud providers, and mechanisms for enumerating bucket
privileges by an unauthenticated attacker are introduced for each
provider.
Worldwide competition, along with a rapidly growing amount of
programs requires hackers to be either very creative in finding unique
bugs or to automate part of their workflow, as only the first hacker
reporting a bug gets paid. Numerous hackers, such as Eric Head, bet-
ter known as todayisnow3 , have proven that the financial gain from a
large scale bug bounty automation may be in hundreds of thousands
of dollars per year [9]. However, most of those tools are kept private,
and there are only very few attempts to release a unified, open-source
automation framework, similar to Metasploit [97] in the penetration
testing field. Further motivation, already existing solutions, and the
technological basis of the proposed framework is discussed in Chap-
ter 4. In Chapter 5, the thesis introduces an attempt to fill this gap and
1. Term “hacker” in this thesis is always used in the meaning of an ethical com-
puter hacker, also referred to as “white-hat hacker.” It is not referring to a person
supporting or doing any illegal activities.
2. National Geospatial-Intelligence Agency (NGA) is a combat support and intelli-
gence agency housed within the U.S. Department of Defense.
3. https://bugcrowd.com/todayisnew
1
1. Introduction
provide a modular bug bounty automation framework, built on top

of modern DevOps systems such as Kubernetes or GitLab.
Results from quantitative research finding Azure Blob Storage
misconfigurations in the wild that revealed numerous potential data
breaches are presented in the Chapter 6. Moreover, based on the
knowledge obtained from both qualitative, and quantitative research
performed, a set of recommendations for object storage providers is
formed and discussed.
2
2 Bug bounty
From a high-level perspective, a bug bounty program is a simple
process – the company publishes a page with assets to get tested, con-
ditions that hackers need to follow and the rewards for the discovery
of valid bugs.
Once a hacker finds a bug, he reports it to the company with a
proof of concept exploit and the possible impact. The company or
a dedicated triage team then validates the report, forwards it to the
relevant team for remediation and rewards the researcher accordingly.
The form of a reward may vary. It may be monetary, in the form of
recognition (public hall of fame) or in the form of branded clothing
(referred to as a “SWAG”).
In the past few years, this concept has proven to be beneficial for
both companies and hackers. Companies receive continuous feedback
on the security of their products, and hackers can earn money legally
by doing what they enjoy, with a company of their choice.
2.1 History
The first known bug bounty program was initiated by Hunter & Ready
in 1983 for their Versatile Real-Time Executive operating system. As a
reward, Volkswagen Beetle (also known as “bug”) or an equivalent
cash reward was offered.
The first modern-day bug bounty program was launched by Netscape
in 1995, offering monetary rewards for the Netscape Navigator 2.0
Beta. Years after, in 2004, Mozilla Foundation started offering boun-
ties for bugs found in their core software. Technological giants like
Google and Facebook followed in 2010 and 2011, respectively. In April
2016, the first Federal bug bounty program, Hack the Pentagon was
launched as a program of a first major, government-backed agency
and the number of programs has substantially increased over the past
few years [1].
Significant growth of programs has opened a vast amount of op-
portunities for hackers around the world. A good example is Santiago
Lopez, who at the age of 19 has earned more than $1,000,000 on the bug
bounty platform HackerOne. By March 2019, he had over 1,683 valid
3
2. Bug bounty
unique vulnerabilities discovered with companies such as Verizon

Media, Twitter or HackerOne itself [9].
2.2 Benefits for the organization

Whether formalized or not, every mature technology company has
a software development lifecycle (SDL). An SDL covers training, re-
quirements, design, development, deployment, testing and response.
While central to the testing and response phases, bug bounty pro-
grams provide valuable insights across all parts of the SDL. A bug
bounty program may point out areas or applications where the com-
pany needs to put more focus or improve overall design practices [10].
For example, it may include severe authentication bugs that would
point out to a missing language-independent authentication across
services in the organization.
However, running a healthy bug bounty program also brings other
benefits like building a strong, security-aware culture across the orga-
nization:
∙ Developers get access to reports from the world best hackers,

which feeds the curiosity about the security industry.
∙ Non-technological departments like business development ac-

knowledge the risks connected with having a public feedback
form, which hackers utilize to submit Blind Cross-Site Scripting
(XSS)1 payloads or similar content injection attacks.
∙ Mid-level management may have an easier time recognizing the

real security risks before an actual breach occurs.
2.3 Comparison with traditional approaches

There are currently two widespread security testing practices – pene-
tration testing and dynamic application security testing (DAST).
1. Blind XSS occurs when the attacker input is saved and then executed in a different
application. This fact makes it extremely dangerous, as the payload may get triggered
in an internal application with privileged accesses.
4
2. Bug bounty
Penetration testing is an authorized process done to find vulnera-

bilities in a given scope and exploit those, in order to make the tested
system more secure [129]. It has been an essential part of the secure
development lifecycle for years, and it is required by certifications such
as The Payment Card Industry Data Security Standard (PCI DSS) [109].
However, as the current trend in development switched, from the wa-
terfall model, where one release cycle could take even six months, to
DevOps that is based on shipping features even multiple times a day,
penetration tests twice a year are not sufficient anymore.
DAST is a program that tries to find vulnerabilities in web appli-
cations, while the application is already running. It performs a black
box test2 by actually performing attacks and analyzing responses of
the application. This automated process can be done very frequently,
but it is unable to cover the application in depth. Penetration testing,
on the other hand, offers a depth of coverage, but is resource intensive
and therefore applied only with a limited frequency. Bug bounties
stand in the middle, covering far more of the application than DAST
and occurring far more frequently than penetration testing [130].
High coverage
Pentest
Bug bounty
Low frequency High frequency
DAST
Low coverage
Figure 2.1: Coverage and the frequency continuance of the different security testing
approaches [130].
2. Black box in this context means that the application does not have access to the
source code.
5
2. Bug bounty
Bug bounty programs are not a replacement for a penetration

testing, as the coverage may not be as in-depth as in a penetration
test and the compliance requirements still need to be met. However,
because of their ongoing nature, bug bounty programs provide a
continuous feedback loop. Combining penetration tests with a breadth
of coverage of bug bounties enables penetration testers to be focused on
the high risk, new, or recently changed application functionality [11].
Penetration testing Bug bounty
Point-in-time – usually one to

Ongoing
two weeks
Single researcher or a small High amount of researchers

team with different skillsets
Tightly scoped – single appli-

Typically broader scope
cation or a specific feature
Mostly performed in staging Performed on production with

environment all defenses enabled
Standards/methodology Results-driven – opens doors

driven – focused on coverage for creative bugs
Paid by time spent Paid per valid bug
Security driven – not a compli-

Usually compliance driven
ance focus
Table 2.1: Comparision of penetration testing and bug bounty [128].
2.4 Self-hosted bug bounty program

Running a bug bounty program can be simplified as having a vulner-
ability disclosure policy (VDP) with details about rewards. A VDP is
intended to give ethical hackers clear guidelines for submitting found
6
2. Bug bounty
security vulnerabilities to organizations [108], but it does not neces-

sarily have to include any mention of a bounty or any commitment on
fixing reported vulnerabilities, only a way how security researchers
can disclose details to the affected company.
Even though the guidelines for writing a VDP got standardized by
the freely accessible ISO/IEC 29147 in 2014 [80], the adoption is still
quite poor. In the middle of 2016, 94% of the Forbes Global 2000 list
did not have the policy to receive, respond, and resolve critical bug
reports submitted by the outside world [87]. At the end of 2017, this
number barely decreased to 93% [1], which shows that this negligence
is more or less stable.
Once a company has a VDP, it takes only a few modifications to turn
it into a full-fledged bug bounty program. This modifications usually
include adding a scope, rules, and rewards. However, companies
need to understand that running a bug bounty program requires
a significant amount of resources, as reports are usually far more
frequent than in case of a VDP. Examples of reputable self-hosted
bug bounty programs are the ones from technological giants like
Facebook [51], or Google [50] that is offering rewards up to $200,000.
2.5 Bug bounty platforms

Another option is to publish a bug bounty program on a Software-
as-a-Service (SaaS) platform, that offers a web application tailored to
the needs of vulnerability disclosure. Some of the benefits that these
platforms offer include an easier payment system, already registered
researchers waiting for new programs and an option for having a team
dedicated to validate and verify the reports, usually known as triagers.
The specific business model of a platform varies from always online
programs to time-restricted tests only lasting a few weeks. As the total
amount of such platforms is now more than 15 [79], only the most
popular platforms will be reviewed.
2.5.1 HackerOne
HackerOne, currently the largest bug bounty platform was founded
in 2012 by Dutch hackers Jobert Abma and Michiel Prins, Head of
7
2. Bug bounty
Product Security at Facebook Alex Rice, and a startup founder Mer-

ijn Terheggen, after frustrating efforts of reporting vulnerabilities to
100 prominent high-tech companies during their challenge called
“Hack 100” [82]. HackerOne quickly gained traction, and in January
2019 it was hosting a bug bounty program for 1,200 companies such
as GitHub, with more than 300,000 hackers registered on the plat-
form [48].
For companies, HackerOne is offering both private and public
programs, in combination with self-managed or fully-managed triage
of reports. In the case of a fully-managed option, HackerOne triage
team is filtering and validating submissions of hackers before they get
to the security team of the customer. Companies usually start with a
private program and few dozen of invited researchers to not overload
the security team or of the application itself. As the program matures,
companies gradually invite more hackers and then potentially make
their program completely public [101].
HackerOne is also very active in community work. At the beginning
of 2018, HackerOne started an open-source educational series aimed
mostly at newcomers to the bug bounty called Hacker101. This series
became a very valuable starting point to new hackers, as it enabled
them to receive invites to private programs just by finishing challenges
from the Hacker101 Capture The Flag (CTF)3 competition [81].
HackerOne also encourages public disclosure of reports after reso-
lution with the feature called Hacktivity [102], that proved to be an
invaluable resource of knowledge for bug bounty hunters and security
teams around the world. Hacktivity, along with a book called Web
Hacking 101 [86] that is based solely on public HackerOne submis-
sions, was one of the main resources for the research on common bugs
in bug bounty done in Chapter 3.
HackerOne ranking system is based on reputation, that the hacker
gains or loses after each report gets closed. The amount of gained
reputation is based on two factors – closing status of the report, such
as resolved or not applicable, and the bounty amount. Total reputation
is then being split into two metrics:
3. Hacker101 CTF is an always online training ground with various intentionally

vulnerable web applications that contain hidden flags for which points are awarded.
8
2. Bug bounty
∙ Signal – average reputation per report. As reputation is gained

or lost each time a report is closed, the signal is an aggregate
representation of report validity. It corresponds to reputation
changes for triage states ranging from “spam” (-10) to “resolved.”
(+7)
∙ Impact – average reputation per bounty. As reputation is gained
based on the relative size of the awarded bounty, Impact is an
aggregate representation of report severity. It corresponds to
reputation gains calculated by bounty levels awarded [103].
2.5.2 Bugcrowd
The business offering of Bugcrowd is roughly the same as HackerOne
except an option for self-managed programs [84]. From the community
perspective, Bugcrowd also introduced an educational series aimed
at beginners called Bugcrowd University. However, it has much less
coverage than Hacker101 at the time of writing, and it does not offer
any option of receiving private invites [83].
The ranking system for hackers is based on “kudos points,” that
are assigned based on the severity of submitted reports. Private invites
are then determined based on the number of points and additional
metrics, such as average severity. Monetary rewards are excluded in
the ranking, which highlights the main difference compared to the
HackerOne ranking system [85].
2.5.3 Others
As previously mentioned, there are many more platforms emerging
in this industry with a similar offering. Examples include:
∙ Hacktrophy – smaller Slovak platform, offering public bug bounty
programs, aimed at the local business. This platform currently
hosts bug bounty program of companies such as ESET, Mall or
Moneta bank [88].
∙ Intigriti – a platform that hosts a majority of programs launched
for open-source software such as Drupal or KeyPass, which
received sponsorship from the European Union [89].
9
3 Common bugs in bug bounty
Bug bounty hunters follow a similar methodology as usual pene-
tration testers, such as OWASP testing guide [104]. One of the most
prevalent vulnerabilities in the world – Cross Site Scripting (XSS) [12,
13, 14], is also the most popular vulnerability among hackers on the
HackerOne [1, 48].
This class of injection bugs has, however, already received sub-
stantial coverage by various researches before [41, 42] and it is also
very rarely present as a root cause for breaches. On the other hand,
two other vulnerability types covered in OWASP Top 10 20171 – Using
Components with Known Vulnerabilities and Security Misconfiguration
were together responsible as a root cause for 44% out of 50 analyzed
breaches in the recent study of Snyk [91].
The focus of this chapter is mostly on similarly impactful, but easily
automatable vulnerability types like security misconfiguration. These
vulnerability types shortly follow XSS in its prevalence, but the dis-
coverability is usually significantly easier to automate [1]. A growing
amount of programs with a broad scope, along with an always-online
nature of a bug bounty programs, make these automatable vulnera-
bility types a perfect candidate for further research.
3.1 Object storage misconfiguration

Object storage is a widely adopted cloud service, mostly used to store
unstructured data like logs, backups, attachments or a static content
of a website. It was designed with high durability and availability in
mind, but the usable security aspect has been considerably overlooked
in the past. Consoles had missing safety nets for less experienced
administrators, that resulted in hundreds of leaks caused by a simple
misconfiguration [132].
All newly created buckets on the tested providers had secure de-
faults, but each cloud provider is taking a slightly different approach
1. The OWASP Top 10 is a report outlining ten most impactful web application
security issues. This report is updated roughly every three years, with 2017 being
the most recent one at the time of writing.
10
3. Common bugs in bug bounty
on how to make its buckets more difficult to misconfigure. I would ar-

gue that Oracle Cloud Object Storage has the least expressive warnings
in UI, along with some less secure default configurations. If not ad-
dressed, these issues could potentially lead to a considerable amount
of future misconfigurations. Recommendations based on the quanti-
tative research performed in the Chapter 6 are further discussed in
Section 6.4 on page 53.
The following areas were analyzed on each provider:
∙ Possible public permissions in the bucket ACLs – a set of pos-

sible permissions which may be assigned to the bucket Access
Control List (ACL) accessible either to an unauthenticated at-
tacker or to an attacker in possession of a dummy account with
the given provider. Each provider allows a different set of per-
missions assignable to the bucket ACL, but it can be generalized
as a subset of the following permissions:
– READ – permission that allows reading all objects by their

full path. Rarely impactful on its own as it requires knowl-
edge of an object name. In AWS S3 terminology, READ
is used for the permission describing LIST in the thesis
terminology.
– LIST – permission that lists the objects in the bucket. If com-
bined with public READ permission, it effectively allows
to dump the content of the bucket. This may have a critical
impact in case the bucket is used to store sensitive data.
– WRITE – permission that allows to create, overwrite, and
delete any object in the bucket. If misconfigured, it often has
a high impact on its own due to the possibility of hosting
malicious files or overwriting valid files, possibly leading
to website takeovers or even remote code execution [47].
– READ_ACP – permission that allows reading the bucket
ACL. The impact is usually low as it only leaks account
names with access to the bucket along with their permis-
sions, but it may be potentially escalated and misused in
some targeted attacks.
11
– WRITE_ACP – permission that gives the ability to modify

the bucket ACL. Overwriting the ACL effectively leads to a
bucket takeover and gives the full control over the bucket to
the attacker. This is the most dangerous misconfiguration,
commonly directly leading to a critical impact.
∙ Ease of misconfiguration – how difficult is it to set public access

on bucket ACL and how well is the impact of the user actions
expressed in the web console. For each provider, screenshot
during or shortly after the bucket creation is included.
∙ Default object ACL – default permissions while uploading a

new object to the bucket.
∙ Enumeration flow of a bucket ACL for an unauthorized at-

tacker – detailed steps which an unauthorized attacker needs
to make to enumerate the bucket ACL. As no previous research
was done on some of the providers, and no modular tool for
object storage misconfiguration analysis was found, I have de-
veloped and open sourced a tool called BucketsPerm [110]. This
tool implements listed workflows for each provider and allows
for easy analysis of bucket permissions at a larger scale.
3.1.1 Amazon Web Services S3

The most popular, and also the most researched object storage service
from Amazon Web Services (AWS) is a well-known source of income in
bug bounty [47]. In 2017, there were 38 times more insecure storage vul-
nerabilities reported on the biggest bug bounty platform HackerOne,
compared to 2016 [1]. AWS S3 has been a noteworthy source of data
breaches for a vast amount of companies worldwide [132, 133].
In the end of 2017 (11 years after the launch of S3), AWS finally
acknowledged the voice of security researchers worldwide, and they
improved the visibility to bucket misconfigurations in the console.
They have added warnings explaining the impact of set permissions,
and also added an orange label for publicly accessible buckets in
the main console [134]. In November 2018, they continued making
improvements with global account bucket controls [135]. Both of those
12
Figure 3.1: Screenshot from the AWS S3 console during the bucket creation.
features helped to improve usable security of S3 dramatically, but there

are likely still many misconfigured S3 buckets left in the wild.
∙ The administrator can assign any set of permissions to an unau-

thenticated user. Each of those misconfigurations except public
READ is possible via a checkbox in the console.
∙ Since November 2018, four different checkboxes (Figure 3.1)

need to be unchecked during bucket creation to allow any future
misconfigurations with a public access [135].
∙ Setting a public bucket policy will trigger a warning with a mes-

sage stating “Everyone has access to one or all of the following:
list objects, write objects, read and write permissions.” I would
argue that this warning is not sufficient and should state the
impact more clearly.
∙ The default ACL for uploaded objects is private regardless of

the bucket policy. This can be either specified during upload
or overwritten by assigning a hand-written READ policy to the
bucket.
Interestingly, it is possible to enumerate existing bucket names only

based on DNS response. This allows for significant optimization of
valid bucket names enumeration and reduces the number of requests
sent to AWS infrastructure to a minimum.
13
DNS CNAME query

<bucket-name>.s3.amazonaws.com
CNAME
else
s3-directional.*
Response
HTTP HEAD to

<bucket-name> Bucket Non-
.s3.amazonaws.com Existent
[200] [403]
HTTP
status code
HTTP PUT with

LIST random path
& data
[200] else
HTTP
status code
HTTP PUT
WRITE
Write ACL
else [200]
HTTP
status code
HTTP GET
WRITE_ACP
Read ACL
[200]
HTTP
status code READ_ACP
Figure 3.2: Enumeration of the S3 buckets.
3.1.2 DigitalOcean Spaces

DigitalOcean, one of the smaller, yet well-known cloud providers
offers its object storage under the name DigitalOcean Spaces. The
API attempts to be fully compatible with AWS S3, so the flow of enu-
meration is very similar except for the bucket identification based on
14
DNS records. Please see Appendix C for the complete enumeration

workflow that was omitted for brevity.
Figure 3.3: Screenshot from the DigitalOcean Spaces console during the bucket
creation.
∙ As DigitalOcean attempts to be fully compatible with the AWS

S3 API, there are the same options of bucket ACL as in the
AWS S3. However, there is an important difference; that only
public LIST permission is possible via the DigitalOcean console.
Other misconfigurations, such as public WRITE or WRITE_ACP
are possible only through the API [93]. Having these settings
mutable only via API adds friction to the process, that arguably
makes it a bit less prone to the misconfiguration.
∙ However, if the administrator assigns policy such as WRITE_ACP

to the bucket ACL through the DigitalOcean API, there is no
indication of this configuration in the console. For example, an
administrator can set a publicly listable policy through the API,
but the file listing is still marked as “restricted” in the console.
As this does not seem as a desired behaviour, it was reported as
a bug to the DigitalOcean through their bug bounty program2
and now awaits their response.
∙ The consequences of having a bucket ACL set as listable are well

explained as seen in Figure C.1, but additional confirmation or
visual elements are missing.
∙ The default object ACL is, similarly as the enumeration workflow

the same as in the case of AWS S3.
2. https://bugcrowd.com/digitalocean
15
3.1.3 Azure Blob Storage

The object storage service from Microsoft Azure, called Azure Blob
Storage, is offering quite a complex set of additional features such
as monitoring or Advanced Threat Protection attempting to detect
malicious activity on buckets. Tests of this feature were not conducted,
but similar anomaly detection algorithms might be useful additional
layer of protection against data leaks caused by misconfigured buckets.
Figure 3.4: Screenshot from the Azure Blob Storage console during the bucket
creation.
∙ It is possible to assign public READ, or READ + LIST policies in

the console, but both are accompanied by a clear warning. Any
other misconfigurations, such as public WRITE policy, are not
possible [92]. There a column dedicated to public access level in
the buckets overview, but further clarification of the impact or
some warning elements are missing.
∙ The uploaded objects are strictly following the bucket ACL [111].
∙ The bucket names are namespaced to the account name, not

global as in case of AWS S3 or Google Cloud Storage. How-
ever, it is possible to enumerate account names based on DNS
responses, as DNS request containing a valid account name
responds with CNAME to blob.<microsoft-id>.store.core.
windows.net. This can be used for effective large scale enumera-
tion of Azure Storage account names.
16
∙ Shared Access Signature (SAS), an equivalent of S3 pre-signed

URLs, is the only option to open a bucket or an object to writing
for unauthenticated users, where the documentation is very
clear about the risks connected with it [112]. There are a few
attack vectors involving SAS URL with writable permissions
remain, but as the discovery is not easily automatable, it is out
of the scope of this research.
DNS CNAME query

<account-name>.blob.core.windows.net
CNAME blob.<microsoft-id>.
store.core.windows.net. else
Response
HTTP GET to Account Non-

/<bucket-name>/random Existent
[ResourceNotFound] [BlobNotFound]
Code
HTTP GET
Bucket Private or
/<bucket-name>?
Non-Existent
restype=container&comp=list
[ResourceNotFound] else
Code
READ READ + LIST
Figure 3.5: Enumeration of the Azure Blob Storage buckets.
3.1.4 Google Cloud Storage

Google Cloud Platform (GCP) offers its Google Cloud Storage (GCS)
service since 2010, yet until recently it was overlooked during bug
hunting efforts. As a critical misconfiguration of GCS is possible in
the console, it may be arguably more prone to misconfiguration. This
17
was recently confirmed by the recently concluded research that dis-

covered 34 critically misconfigured buckets among Alexa Top 10,000
websites [136].
Figure 3.6: Screenshot from the Google Cloud Storage console after the creation of a
public bucket.
∙ All misconfiguration options are possible, but GCS console does

not provide a simple checkbox in the console to make the bucket
publicly accessible. The administrator needs to write the name
of a special group allAuthenticatedUsers (everyone with Google
account) or allUsers (unauthenticated users) while assigning a
policy to the bucket.
∙ During the misconfiguration, there is no warning explaining the
possible impact displayed to the user. However, the GCS console
starts displaying a permanent warning explaining the risks right
after the bucket is misconfigured.
∙ The objects can be forced to inherit the bucket ACL by enabling
a new beta feature called Bucket Policy Only (similarly to AWS
S3).
∙ Privilege enumeration is possible via the testPermissions end-
point3 instead of having to check each permission individu-
ally like in the rest of tested providers. Two requests should be
made: one unauthenticated (group allUsers), and the second one
authenticated with any Google account (group allAuthenticate-
dUsers).
3. https://www.googleapis.com/storage/v1/b/<bucket-name>/iam/
testPermissions
18
HTTP HEAD to

googleapis.com/storage/v1/b/<bucket-name>
else [400 || 404]

HTTP
status code
HTTP GET
/iam/testPermissions
Bucket
READ LIST WRITE WRITE_ACP
Non-Existent
Figure 3.7: Enumeration of the Google Cloud Storage buckets.
3.1.5 Alibaba Object Storage Service

Alibaba Cloud, also known as Aliyun (aliyun.com), is a Chinese cloud
computing company that offers its object storage service under the
name Alibaba Object Storage Service (OSS).
Figure 3.8: Screenshot from the Alibaba Object Storage Service console during the
bucket creation.
∙ There are only two options for public misconfiguration – READ

(Public Read) and READ + WRITE (Public Read/Write) policy,
both accessible directly from the console. No option to assign
public LIST permission to the bucket was found, which is unique
among the tested providers and greatly limits the exploitability
of the remaining misconfigurations.
∙ Both, Public Read and Public Read/Write, are accompanied by
a clear warning and a prompt asking for confirmation.
∙ The default setting for a new object is set to inherit the ACL from
the bucket, but it can be changed on an object level, similarly to
AWS S3.
19
HTTP GET to

<bucket-name>.oss-eu-west-1.aliyuncs.com
[AccessDenied] [NoSuchBucket]
Code
HTTP GET with [NoSuchKey]

Bucket
region
Non-Existent
from <Endpoint>
[Anonymous user has no right [The specified key
to access this bucket.] does not exist.]
Message
HTTP PUT with

Private random path &
data
[200] [403]
HTTP
READ + WRITE status code READ
Figure 3.9: Enumeration of the Alibaba OSS buckets.
3.1.6 Oracle Cloud Object Storage

Oracle is one of the major players on the market, differentiating mostly
with its data-as-a-service offering. From the usability perspective,
the console feels the least user-friendly from all of the tested cloud
providers.
Figure 3.10: Screenshot from the Oracle Cloud Object Storage console while updating
the visibility.
20
∙ Oracle allows only a limited set of public permissions – READ

or READ + LIST policy. Both options are accessible through the
console.
∙ Compared to the rest of the tested cloud providers, Oracle pro-
vides the least expressive warning about the consequences of
public listable access. It is missing visual elements and, more
importantly, a prompt with clearly explained impact.
∙ Insecure defaults are present in case the user decides to change
the bucket from private to public. The checkbox “Allow users
to list objects from this bucket” is turned on by default, that
effectively makes a rarely suitable, and dangerous LIST policy,
the default option for the public buckets.
∙ On the other hand, the enumeration of Oracle Cloud Object
Storage ACL is the most difficult among the tested providers due
to the very generic error messages. There is no way to enumerate
valid namespaces or to know if a correct region was found, that
greatly increases the difficulty of the enumeration process itself.
∙ Default setting for new objects is to inherit ACL from the bucket.
HTTP GET to

<namespace>.compat.objectstorage.
<region>.oraclecloud.com/<bucket-name>/test
[NoSuchKey] [NoSuchBucket]
Code
HTTP GET Private or Non-

.../<bucket-name>/ Existent
[NoSuchBucket] else
Code
READ READ + LIST
Figure 3.11: Enumeration of the Oracle Cloud Object Storage buckets.
21
3.2 Leaked sensitive API keys

Leaked credentials were always a critical issue that with the rise of
cloud services only increased in its prevalence and severity. API keys
to sensitive environments like AWS, Github, Slack or GCP are being
leaked and exploited every day [16, 17, 15]. The most common sources
are:
∙ Git repositories – leaked credentials are sometimes present in

the organization repositories, or more frequently in one of the
personal repositories of an employee [17, 106]. Extensive aca-
demic research on secret keys leaked in GitHub repositories by
researchers from North Carolina State University yielded alarm-
ing results. Every day, thousands of new API keys are being
pushed into public GitHub repositories, with more than 100,000
currently affected repositories overall [7].
∙ Logs of continuous integration (CI) services – sometimes com-

panies either misconfigure their CI/CD service (such as Jenkins)
to allow access from unauthorized users that then get access to
the logs, or private tokens leak in intentionally public CI logs of
various open-source projects [18, 107].
∙ Embeded in application – mobile or electron application devel-

opers tend to forget that everything they embed in the client-side
application will be accessible by anyone in possession of the bi-
nary [15]. Sometimes, developers are also relying on client-side
obfuscation, not realizing the fact that any client-side obfusca-
tion can be deobfuscated.
∙ Forgotten files on web servers – it may happen that during

debugging or doing some maintenance work, some file gets
included on the web server. Impact of it may vary from mostly
harmless .DS_Store files4 , up to the full access to the organization
repository. A great example from this category includes a cloned
private repository of Slack being temporarily available on one
4. .DS_Store is Apple OS X specific file used to store attributes of folders with paths
as described in the report https://hackerone.com/reports/142549
22
of the quality assurance servers under the /.git/config path for a

short period of time [128].
Continuous monitoring is vital in case of similar “ephemeral” vul-

nerabilities, especially nowadays when both GitHub [20] and GitLab
is searching for secrets in the code. This monitoring can be achieved by
periodically running numerous open-source tools like TruffleHog [21]
or gitleaks [22], searching for high-entropy strings or strings matching
specific regular expressions. However, the research “How Bad Can it
Git?” mentioned above, showed that even the best open-source tools
such as TruffleHog are detecting only 29% of secrets compared to their
methods [7].
3.3 Broken Authentication

Missing authentication or insufficient authorization is a widespread
critical problem. In the OWASP Top 10 – 2017 list, Broken Authenti-
cation vulnerability type placed second, and Broken Access Control
was on fifth place [14]. In bug bounties, it is a similarly prevalent issue.
With the DevOps culture and growing amount of assets, companies
sometimes deploy services publicly, and they forget to add an au-
thentication layer on top of it, as open-source tools frequently do not
have a built-in authentication. This leads to unauthenticated services
revealing sensitive information or providing destructive functionality.
Examples of commonly exposed services:
∙ Jenkins – the impact of misconfigured authentication on Jenkins

CI servers may lead from the unintentionally visible CI logs
revealing sensitive data, up to Remote Code Execution using its
plugins [18, 24, 23].
∙ Kibana / Elasticsearch – Elasticsearch, Logstash, Kibana stack

(ELK) is widely deployed in various companies. However, the
user interface (Kibana) does not have an authentication mecha-
nism built-in in the basic version [105]. This frequently leads to
publicly accessible Kibana dashboards, whose impact heavily
varies based on the nature of data stored inside of the Elastic-
search cluster [28, 25, 27].
23
∙ MongoDB – insecure default configuration of MongoDB con-

nected with its popularity caused havoc among the noticeable
amount of companies hit by ransomware attacks, that encrypted
their database. Even after several years, the number of miscon-
figured MongoDB instances in the wild is still alarming, and it
continues to be a common source of data breaches [29].
∙ Django Admin – admin login typically present on the /admin

endpoint on the web server is often missing any password brute-
force protection, that many times allows to guess the correct
admin credentials. Once the attacker gains access to the Django
admin site, he can usually effectively compromise the website
or the underlying database [34].
The automated discovery of similar bugs is usually achieved by
subdomain enumeration tools discovering new assets. These assets
are then periodically fingerprinted and, in the case of web services,
a screenshot is taken. In case of any change, a notification is sent for
further manual analysis, as some custom services cannot be possibly
covered only by fingerprinting.
3.4 Security Misconfiguration

This type of vulnerability, representing the sixth place in OWASP Top
10 – 2017 [14], covers similar issues as previously discussed authen-
tication errors. Some services still suffer from insecure defaults, or
administrators enable debug features for testing and then forget to
disable them on production. Severity is usually not as high as with
previously mentioned categories, but its prevalence is compensating
this fact in average bounties. Examples include:
∙ Exposed phpinfo – very common misconfiguration where PHP
application leaves this debug functionality enabled on produc-
tion. This page then allows anyone to request detailed informa-
tion about the system on which the application is running [31,
32].
∙ Django debug mode – leaving debug mode turned on Python

Django framework may result in various problems, the most
24
evident is that in case of exception in program flow, Django

will display a detailed traceback, including a lot of metadata
about application environment, such as all the currently defined
Django settings. These settings usually contain secret tokens
and configuration details, that, if exposed, may have devastating
consequences. Documentation is unambiguous, stating “Never
deploy a site into production with DEBUG turned on” [33]. This
specific vulnerability was once a root cause of Remote Code
Execution on Facebook servers that led to a reward of $5,000 in
their bug bounty program [35].
∙ Directory listing – it is common to leave directory listing en-

abled on a web server such as Apache which has this feature
enabled by default. This problem is prevalent in various Word-
Press installations, potentially leading to the exposure of some
sensitive files [36].
Similarly, as in the previous category, the discovery of security

misconfiguration may be done by various vulnerability scanners with
fingerprints for known vulnerabilities. Examples of open-source tools
are Snallygaster [45] or Inception [44] that attempts to spread the load
evenly across a vast amount of hosts. As both tools lacked necessary
flexibility for my needs, I wrote and open-sourced a tool called Low-
hanging [46].
3.5 Subdomain takeover

With the adoption of DevOps approach, digital footprint5 of numerous
companies IS currently increasing. Previously internal applications
are often getting a public DNS record, that, along with an inefficient
management of domain names, may frequently cause dangling DNS
records6 pointing to various cloud resources such as an AWS S3 bucket
or Heroku application [126, 127]. Once claimed on the side of the
resource provider, attacker can get full control over the subdomain
5. Cloud assets, DNS records, third-party services, and so forth

6. DNS record is considered dangling if the resource to which it is pointing to can
be manipulated by a third party other than the one who controls the record.
25
and misuse it to launch phishing attacks, or gain access to the cookies

of the parent domain.
Automation of this vulnerability class can be achieved by a con-
tinuous monitoring of DNS records, periodically sending an HTTP
request and comparing the output against a list of fingerprints with
applications susceptible to subdomain takeover. There are numer-
ous open source tools helping with discovery, such as Subjack [122]
or Tko-subs [121]. Furthermore, there is a repository which attempts
to centralize knowledge about subdomain takeovers, along with its
fingerprints and exploitability named can-i-take-over-xyz [120].
3.6 Server Side Request Forgery

Even though Server Side Request Forgery (SSRF) is not easily de-
tectable by automated scanners, the severity is rising dramatically
with the growing adoption of cloud services. In connection with in-
stance metadata server on several cloud providers like AWS or Google
Cloud Platform, it sometimes allows for escalating up to the full com-
promise of the company account [37]. The instance metadata server is
a special server present on 169.254.169.254, used to store details such
as hostname, zone, tags, associated SSH keys and cloud provider keys.
For reads and writes from within an instance, the metadata server pro-
vides automatic instance-level authentication and authorization. Each
instance can read or write only to its own metadata server [38, 40].
In the SSRF attack, the attacker can abuse functionality on the
server to request internal resources originating from the host itself.
Example of such internal resource is the instance metadata server
on AWS, that may then reveal AWS API tokens. In case the IAM per-
missions7 are lax, and the attacker can retrieve keys belonging to the
instance, he may completely take over the AWS account of the victim.
Google Cloud Platform recently added protections to its meta-
data server to address this issue. Every request to metadata server
now must contain Metadata-Flavor: Google header and every request
7. AWS Identity and Access Management (IAM) is a web service that helps securely
control access to AWS resources. It may be used to assign granular permissions to
each instance or service, but it is commonly left with an unnecessarily permissive
policy by AWS administrators.
26
with X-Forwarded-For header is discarded as it generally indicates that

the requests are proxied. Both of these protections are, however, still
bypassable through legacy v1beta1 metadata endpoint [38], so the
functionality. AWS, as the most popular cloud provider [131] does not
take similar precautions yet and leaves securing of instance metadata
server as a responsibility of the administrator.
The intriguing fact is that SSRF is currently on a relatively low
rank in the OWASP Top 10 project on the 28th place as based on the
dataset used in 2017, only 1.69% out of 114,000 applications contain
this vulnerability [39]. In the most recent HackerOne Top 15 list, this
vulnerability was, however, ranked on 15th spot [67], that may even-
tually forecast its possible rise in the upcoming years.
27
4 Bug bounty automation
Bug bounty automation is currently a subject undergoing intense
study in the bug bounty community. In this chapter, the motivation
for working on such automation, along with the requirements for a
potential bug bounty automation framework will be discussed. Some
existing solutions that emerged during the past two years will be
discussed, but as none of the presented solutions fulfills the set of
requirements, the technological basis of the proposed solution will be
introduced.
4.1 Motivation
Bug bounty, being a very young and competitive scene, lacks a com-
mon platform for automation of tedious tasks. Various open-source
command line tools are released frequently, but there is currently no
common framework that would allow combining them easily. Even if
such a framework is created, it is either kept private or shared only in
small groups with the indisputable motivation of receiving a bounty.
A robust, scalable and easily extensible solution is missing, and many
talented bug hunters around the globe are forced to reimplement it
over and over again. This goes against a known open-source principle
that “no problem should ever have to be solved twice,” also mentioned
in the How To Become A Hacker article by Eric Steven Raymond [99].
Existing penetration testing frameworks like Metasploit [97] may
be useful, but they were initially made for a different use-case. These
tools are only semi-automatic, and they mostly lack continuous moni-
toring functionality. In the end, they were done a different use case,
as a penetration testing is usually done in depth on a narrow scope
during a restricted timeframe.
Conversely, DAST, that is a fully automatic solution capable of
continuous monitoring, was not built to provide a base ground for
further manual testing and only focuses on finding vulnerabilities
based on specific fingerprints. Another disadvantage of DAST is that
it usually causes a considerable amount of traffic that may potentially
overwhelm the servers of the target.
28
4. Bug bounty automation
Sitting in the middle, an ideal bug bounty tool should arguably

satisfy the following requirements:
∙ Scalable – the number of targets should be able to scale up to

hundreds of wildcard domains, to be able to monitor all public
bug bounty programs simultaneously.
∙ Modular – bug bounty techniques and tools are advancing

quickly, so adding new functionality has to be straightforward.
∙ Segregated duties and resources – every step should have clearly

defined input, output, separated logs, and resource usage. A sin-
gle misbehaving step should not be able to impact the rest of the
running tasks negatively.
∙ Provide a base for manual testing – results from the tool should
be able to serve as a good resource for further manual analysis
or testing.
∙ Open-source – as the framework is mostly using already created

open-source tools; it should also be open-source.
4.2 Existing solutions

I would argue that providing a robust framework should be in the inter-
est of bug bounty platforms. This would benefit both sides as hackers
would not have to spend less time on automation that could conse-
quently result in an increased amount of valuable, manual findings
that benefit the platform. However, none of the platforms currently
attempts to solve this issue. Some of the tools from bug bounty com-
munity emerged, but as the community is quite decentralized, they
are mostly made by a single contributor.
∙ Bounty Machine – a tool that inspired the core infrastructure

of later proposed solution. Built on top of Argo, with queues
and proper database, this tool seems like a complete product.
The original plan of this research was to contribute and help
develop this already existing tool, but unfortunately, authors
have decided to keep this tool closed source [94].
29
∙ Intrigue Core – fairly sophisticated tool written in Ruby with

dozens of tests already included. Intrigue Core abstracts the
attack surface into graphs where nodes represent an asset and
edge are a task. This tool also comes with a built-in web inter-
face, from which tasks can be initialized and results analyzed.
A disadvantage of this solution is that the application itself is a
monolithic ruby app, and thus, a single misbehaving tool can
take down the whole application. Because of its monolithic na-
ture, extensibility of this solution outside of native Ruby code
may also be a problem as system changes could potentially af-
fect dependencies of other already implemented tools [95]. The
proposed solution may end up integrating Intrigue Core as one
of the modules in the future.
∙ Lazy Recon – series of bash scripts by various authors such

as Ben Sadeghipour that offers a quick and dirty solution for
bug bounty automation. However, due to its bash nature, re-
quirements such as modularity, compartmentalization or future
scalability are not realistic [96].
As no existing solution matching all specified requirements was

found, the rest of the thesis will introduce the proposed solution that
was developed as an attempt to fill this gap. This proposed tool will
be further referred to as Bugshop.
4.3 Technology used
All technologies used in the proposed framework are currently being

utilized in the world leading technology companies such as Google.
Interestingly, the utilization of modern DevOps technologies such as
Docker is still rare in the security community. During the development
of Bugshop, I have contributed Dockerfiles with an example of docker
usage to some popular offensive security tools, such as S3Scanner [113]
or Altdns [114], that positively impacted awareness about the benefits
of Docker in the security community.
30
4.3.1 Docker
Docker is a tool released in 2013, designed to make it easier to create,

deploy, and run applications by using containers. Containers allow
developers to package up an application with all underlying parts it
needs, such as libraries and other dependencies and ship it out as a
single package. By doing so, it is assured that the application will run
in the same environment on any other machine.
In a way, Docker is a bit like a virtual machine. However, unlike
a virtual machine, Docker allows applications to use the same Linux
kernel as the system that they are running on and only requires appli-
cations be shipped with things not running on the host computer. This
gives docker a significant performance boost while reducing the size
of the application [59]. Docker itself is, by now, a quite well-researched
tool with various academic studies on its usability and security [60, 61],
where further recommendations can be found.
Bugshop requires dockerized applications due to its underlying
Kubernetes native infrastructure. However, “dockerization” also effec-
tively abstracts the language and package requirements and enables
to create an optimal environment for any given tool. The nature of
containers also helps with debugging, as the production environment
effectively matches the one on the local machine and real-time metrics
about CPU usage, memory, consumed disk and network bandwidth,
are available using docker stats [116]. Furthermore, Docker assures
that a single misbehaving tool cannot crash the whole cluster, as it is
possible to apply CPU and memory limits on each container easily.
4.3.2 Kubernetes
Kubernetes is an open-source system for automating deployment,

scaling, and management of containerized applications (usually, but
not limited to Docker). Kubernetes is decoupling the application con-
tainers from the details of the system, and it also resolves networking
within the cluster. It was built upon 15 years of experience of run-
ning production workloads at Google, and it allows scaling of a single
cluster up to 5000 nodes (hosts) and 300,000 containers [60, 62, 63].
Choice of this technology was constrained the by choice of Argo, but
31
Kubernetes has lately established itself as the de facto standard for

container orchestration [64].
Kubernetes is providing or supporting every functionality of Argo
and provides easy scalability and deployment. During my develop-
ment, I have used managed Kubernetes on Google Cloud Platform
(GCP) called Google Kubernetes Engine (GKE). GCP offers a trial
period for new accounts with free $300 credits can be spent over one
year [65]. Thanks to this, I was able to test and develop my solution
without spending anything on the infrastructure. Their managed so-
lution also effectively reduced the deployment time of Kubernetes to
zero as the GKE comes preconfigured with everything required by
Argo. Google also offers some convenient functionality like real-time
monitoring of resource usage per container (as seen in Figure 4.1)
and Cloud Identity-Aware Proxy [66] that was utilized to provide
authentication to the web interface of Argo.
Figure 4.1: Resource usage graph of a container in the GKE console.
4.3.3 Argoproj
Argoproj is a collection of open-source tools for getting work done
with Kubernetes. While being a very young project where the first
32
stable version 1.0.0 was released in August 2017, it is already being

used by companies such as Google, NVIDIA, Datadog or Adobe for a
part of their workload [68]. In addition to that, some popular open-
source tools for Machine Learning, such as Kubeflow Pipelines were
already built on top of Argo [69].
The original project of Argoproj is called Argo Workflows (also
referred to as Argo), but by now, it consists of several parts:
∙ Argo Workflows – container-native Workflow Engine (utilized

in Bugshop).
∙ Argo Events – event-based Dependency Manager (utilized in

Bugshop).
∙ Argo CD – declarative GitOps Continuous Delivery.
∙ Argo CI – simple CI based on GitHub and Argo Workflows

(currently inactive).
Alternatives of Argo include Apache Airflow [72] and Brigade made

by Microsoft [73]. After a short evaluation, Argo was chosen due to
its simplicity and active, helpful community built around it.
Argo Workflows
Argo Workflows is an open-source container-native workflow engine
for orchestrating parallel jobs on Kubernetes. Argo Workflows are
implemented as a Kubernetes Custom Resource Definition (CRD).
It offers to define workflows where each step in the workflow is a
separate Docker container with clearly defined responsibilities.
These steps can be then chained into multi-step workflows, where
each level consists of one or more steps (containers). An alternative
approach can capture the dependencies between tasks using a di-
rected acyclic graph (DAG), similar to Apache Airflow. The multi-step
workflow declaration is more straightforward to define and allows an
option of dynamic workflows, where DAG based declaration offers
much more granularity needed for the optimization of more complex
workflows, such as buckets enumeration [71]. For the screenshot of
bucket workflow, please see Figure 5.4 on page 46.
33
Workflows are represented by a yaml configuration file, where

steps with clearly defined input, output, docker image, and artifacts.
Example snippet of a single step can be seen on listing 4.1. This en-
ables easy debugging and high level of modularity – if a new workflow
needs to be developed, separate workflow file is created and deployed
without potentially affecting any of the existing workflows. A signif-
icant advantage of this segregation is also a separation of logs and
monitoring of resource usage, thanks to which the failure point is
easily identified and debugged. In Bugshop, this is an essential feature
as it mostly utilizes third-party open-source tools with questionable
reliability.
− name : amass
container :
image : s14ve/amass : 2 . 9 . 9
volumeMounts :
− name : c o n f i g
mountPath : "/ c o n f i g "
resources :
requests :
memory : 4Gi
args :
[
"−d " , " { { workflow . parameters . domain } } " ,
"− c " , "/ c o n f i g /amass . i n i " ,
"− o " , "/tmp/amass . t x t "
]
outputs :
artifacts :
− name : r e s u l t s
path : /tmp/amass . t x t
Listing 4.1: Confuration snippet of a single step in subdomain enu-
meration workflow.
Argo Events
Argo Events is an event-based dependency manager for Kubernetes
that allows defining multiple dependencies from a variety of event
34
sources like webhooks, S3 notifications or calendar schedules. After

the successful resolution of specified event dependencies, it triggers
specified Argo Workflows. Events also offer additional functionality
such as defining arbitrary logic after the resolution of dependencies
or filtering of inputs, but it was not utilized for a simple use case of
Bugshop.
Argo Events is made up of two main components, both imple-
mented as a Kubernetes Custom Resource Definition:
∙ Gateway that processes events from the event sources, such as
HTTP webhook or S3 event notifications.
∙ Sensor that defines a set of event dependencies and after its ful-
fillment triggers Kubernetes resources, such as Argo Workflow.
4.3.4 GitLab
GitLab is a web application that provides a Git-repository manager
with wiki, issue-tracker and a native CI/CD pipelines for free [75]. In
Bugshop, GitLab with its CI/CD is utilized not only as a repository
of the source code but also as a crontab scheduler and storage of
workflow results.
Commit history allows for easy tracking of changes (see Figure 4.2)
and various integrations such as built-in alerting on changes can be
further utilized. Using Git repository as a database in bug bounty was
inspired by a blog post about the bug bounty monitoring [76] by the
author of security.txt IETF draft [77] – Edwin Foudil. GitLab may get
replaced by a full-fledged database in the future, but it is sufficient for
the current requirements of Bugshop.
Figure 4.2: View of the changes on a commit in GitLab.
35
5 Proposed solution
As none of the existing solutions matched all specified requirements,
a new automation framework was implemented with emphasis on
easy extendibility and high scalability. The aim of this solution was to
abstract the underlying technology and provide a reliable, modular
framework for any bug bounty hunter with an intermediate technical
background. The proposed solution can be deployed only with an
elementary knowledge of Kubernetes and extended easily by adding
or editing existing workflows via definitions written in YAML serial-
ization language. The framework is currently split into two standalone
components:
∙ Bugshop – the core of the tool that includes workflow definitions,
configuration files, and Kubernetes manifests used to deploy
Argo and Argo Events in the desired state.
∙ Bugtab – short python program and set of configuration files

used to schedule tasks via GitLab CI pipeline schedules.
Majority of the Bugshop functionality comes from the utilization
of already existing tools such as Argo or GitLab. Connecting existing
tools to create something bigger instead of reimplementing it from
the grounds up is ultimately the philosophy that Bugshop tries to em-
brace, similarly to UNIX philosophy of making each program do one
thing well [115]. During the implementation of specific workflows,
some gaps in open-source tooling have been identified, for which
python tools following UNIX philosophy were developed and pub-
lished under my GitHub account1 . Some of them will be introduced
in Section 5.3, dedicated to the implementation of workflows.
5.1 Architecture
Firstly, the terminology used in further sections needs to be defined.
Trigger is used to describe an initiator of the task. It has a form of an
HTTP request. Notification is a trigger initiated by some action, such
1. https://github.com/janmasarik
36
5. Proposed solution
as detection of a new public program or a successfully finished task,

where task is an instance of workflow with defined input parameters,
such as domain name. Workflow is a set of tools, characterized as
docker images in a specific order. For a more detailed explanation,
please refer to Section 4.3.3. Record is a file with a specific structure
that usually represents an output of a single workflow.
From a high-level perspective, the architecture of Bugshop can be
split into four main phases:
1. Trigger – each new task is initiated by a trigger. Sources of trig-

gers vary, but they can be generally split into three categories –
invoked periodically, by notification, or manually by an operator
that engages a new target. Triggering a new task is currently
possible either via Argo Events HTTP gateway or via Kubernetes
API.
2. Process – once a task is triggered, it starts the selected workflow

and parses the results into some machine-readable output. Op-
tionally, a task may trigger a notification to start new, normally
dependent tasks. Processing of tasks is handled by the Argo
Workflows engine in the Kubernetes cluster.
3. Store – parsed results of task are being stored in a persistent

database with a log of past changes for further analysis. Bugshop
uses GitLab repositories with a defined file structure as a storage
for the results of tasks, and raw output of each step of every task
is saved in a dedicated Google Cloud Storage bucket.
4. Alert – a change in some of the records usually means a new

asset or a newly detected vulnerability. Bugshop alerts the oper-
ator on a change in selected files so he can evaluate the change
in real time. Alerting is provided by GitLab’s Slack or email inte-
grations that can be set up to send a message on each push along
with an overview of changes. (see Figure 5.6). In the future, this
step will be partially replaced by an automatic submission of
selected findings.
37
1. Trigger 2. Process 3. Store
GitLab CI
Docker GitLab
Argo Events
HTTP Gateway
Slack
Email
Argo Slack
Kubernetes API Workflows Kubernetes Notification
Argo CLI Cluster 4. Alert
Figure 5.1: High level overview of the Bugshop.
5.2 Triggers
As shown in Figure 5.1, there are currently two trigger listeners imple-
mented in BugShop; Argo Events gateway and Kubernetes API. Argo
Events HTTP gateway is utilized by Bugtab, Slack webhook or as a
listener for notifications from the onExit handler of workflows. New
triggers can be added easily, as the gateway is essentially just a web
server listening for POST requests with the specified JSON payload.
Listener for Kuberentes API is used only via the Argo command line
interface (Argo CLI), as it requires authentication to the Kubernetes
cluster.
5.2.1 Manual
The simplest way how to trigger a workflow is by the Argo CLI. This
command line interface from Argoproj offers commands for workflow
initialization, termination, status check or logs. Argo CLI is essentially
just a wrapper around the Kubernetes command line tool called kubectl,
that allows you to run commands against Kubernetes clusters [52].
The additional value of Argo CLI is in features like syntax validation
of workflow manifests, more readable output, and more intuitive
interface. For example, the following command would be used to
trigger a new workflow for subdomain enumeration:
argo submit subdomains-workflow.yml -p domain=example.com
38
5.2.2 Bugtab
Continuous monitoring is a fundamental feature of Bugshop that was
originally planned to be implemented via the Argo Events calendar
gateway. However, implementing it this way would require too many
lines of duplicate code for each new target, which was a significant
drawback considering the need for an easy extendibility of the scope.
Furthermore, other use-cases requiring custom code emerged, such
as automatically starting a full monitoring string on both, new public
programs and new invites to private programs on HackerOne.
Because of this, a python utility called Bugtab was implemented,
where GitLab CI schedules are then being used to execute it periodi-
cally and keep track of time. The architecture of Bugtab, represented
as Gitlab CI in Figure 5.1 will be further described in this subsection
and can be seen in detail in Figure 5.2.
Thanks to Bugtab, it was possible to achieve an easy extendibility
of scope, as starting a continuous monitoring process on a new target
requires only a few added lines into the scope, as illustrated in the
Listing 5.1 below, and it also saves time by automatic monitoring of
new bug bounty programs. Bugtab is designed in a modular way
to allow others to potentially make use of it without having to use
Bugshop or Gitlab CI.
Listing 5.1: Example entry in scope.yaml

starbucks :
domains :
− s t a r b u c k s . com
− teavana . com
schedule : a g g r e s s i v e
Schedule
Bugtab currently consists of two configuration files config.yaml and
scope.yaml. These two files are being parsed and linked together by
Bugtab that initiates a series of HTTP requests containing the name
of workflow along with required parameters to the HTTP gateway
of Argo Events. To keep resulting configuration short, it groups the
sets of modules into so-called “schedules” that contain details on how
39
config.yml
single task
bounty-targets-data Schedule Argo
Workflows
chained task
New public Argo Events

Google Cloud
program detection HTTP Gateway
Storage
Kubernetes
Cluster
HackerOne HackerOne
GraphQL API private invites
detection GitLab CI
Figure 5.2: Detailed overview of the Bugtab.
often should each module run. Schedules are then assigned to each
target, as seen on Listing 5.1. Currently, there are only two schedules
implemented – aggressive and light, however, adding a new schedule
takes just a few lines of configuration, as seen on Listing 5.2.
Listing 5.2: Example configuration of schedule in config.yaml

aggressive :
− module : subdomains . yml
frequency : 4h
− module : subtakeover . yml
frequency : 30m
− module : b u c k e t s . yml
frequency : 1d
To achieve a straightforward implementation, new GitLab CI pipeline

schedule is set up for each desired interval, such as 30 minutes. It then
passes the frequency identifier to the script that takes all modules that
are supposed to run, parses the scope.yaml file and triggers the tasks.
40
New public program

Besides schedules, Bugtab also implements a convenient feature that
monitors a list of public programs of the two major bug bounty plat-
forms – HackerOne and Bugcrowd. Once it detects a new wildcard
domain to scan, it starts a special chained task that will trigger all
available workflows in the relevant order.
This is implemented by hourly monitoring of GitHub repository
bounty-targets-data that contains all wildcard domains from the scopes
of public bug bounty programs. Each run, the program fetches the
current wildcards.txt file, fetches the wildcard.txt file from the pre-
vious run saved in GCS bucket, and performs a set subtraction. If
any changes were detected, it triggers an API request to Argo Events
gateway. Afterward, it saves the current wildcards.txt file in the GCS
bucket to prevent redundant executions in the next run.
Private program invite

To add an additional level of automation, Bugtab is also triggering
a scan upon receiving a private program invitation. In 2017, 79% of
bug bounty programs on HackerOne platform were private, invite-
only programs with only dozens of hackers participating [1]. It is
usually easier to find bugs in private programs, as there is generally
fewer hackers actively participating in it that makes private program
invitations a valuable resource.
To get an invitation, a hacker needs to have a certain reputation,
non-negative signal, and some activity in recent three months. The
number of invites one hacker can get is limited, but if a hacker declines
an invite on HackerOne and provides a reason why he did so, he
is automatically assigned to the priority queue for new invitations.
Priority queue effectively means that a new invite for the different
program is sent the day after [117]. This causes hackers to choose
programs in which they are going to participate very carefully.
As the new invitations are on HackerOne usually sent around
noon, Gitlab CI schedule set for 13:00 checks for invites received in
past 24 hours and potentially triggers a scan. Triggering a fully fledged
scan on received program invitation allows the operator to make an
informed decision faster, as additional information is available during
41
the decision process. This functionality is not finished in the Bugtab

as of the time writing this thesis (May 19).
5.3 Workflows
The need for some workflow engine emerged while trying to do more
complex, chained tasks, such as general bucket enumeration. There
are numerous open-source, offensive bug bounty tools available and
they are being grouped to various directories such as the one on Bug
Bounty Forum2 . However, new tools are often being created instead of
chaining already existing tools together. A good example is mass33 that
was created for enumeration of S3 buckets on the DNS level. There
is, however, already an existing tool for brute forcing DNS called
MassDNS4 .Instead of implementing a new tool, it is easy to chain
MassDNS to brute-force DNS lookups, sed to filter only valid buckets
names, and then optionally run BucketsPerm to audit the permissions
of the buckets found.
5.3.1 Subdomain enumeration

When the bug bounty hunter first approaches a target with wildcard
domain in scope, the first thing that he usually does is to run some sub-
domain enumeration tool. Various tools for this use case has emerged
in the past five years. Examples of well maintained tools from this
category are Subfinder or Amass. Both of these tools utilize various pas-
sive or active techniques for gathering subdomains, such as Shodan,
Censys, Webarchive or Certificate Transparency.
These tools have some unique data sources and usually provide
a slightly different set of results. Instead of choosing a single tool
and relying on the maintainers to keep the sources up to date, we
can instead just use them both, add MassDNS for brute-forcing and
then merge the results. Merge process is done by a tool call Resolvable,
during which the results are sorted, validated, and enriched by some
additional information such as AAAA DNS records. This decreases
2. https://bugbountyforum.com/tools
3. https://github.com/smiegles/mass3
4. https://github.com/blechschmidt/massdns
42
the maintenance burden as even if one tool starts malfunctioning, the

workflow will continue to work.
In the implemented workflow, this process is further extended
by AltDNS that performs alternations of the existing subdomains
in order to find similarly named domains, such as subdomain-dev.
example.com from subdomain.example.com. This list is then passed
to MassDNS, and the resulting list is again merged, resolved, enriched
and stored.
Root domain
Subfinder Amass MassDNS
Resolvable parse MassDNS
AltDNS MassDNS
Resolvable parse MassDNS
Enriched
Sudomains resolvable
subdomains
Figure 5.3: Subdomain enumeration workflow implemented in the Bugshop.
5.3.2 Bucket enumeration

Insecure buckets which got great coverage in the Section 3.1, are still
a prevalent problem. This workflow tried to address it and brings
some additional steps to the usual bucket enumeration workflows.
43
This workflow is implemented using the DAG model, as because of

the complexity of the workflow, specifying granular dependancies
moderately improved the performance. Part of this workflow can be
seen in Figure 5.4.
The usual process of bucket enumeration takes a company name
and generates a list of possible candidates by doing permutations with
some wordlist. For example, if a company would be named acme.com,
the usual approach would generate bucket names such as acme-dev,
acme-assets, or acme-attachments. This approach proved to be quite
effective, but as misconfiguration of S3 buckets is now regularly tested
by a vast amount of researchers, most of the similarly named miscon-
figured buckets were already found and fixed. As AWS S3 considerably
improved the usable security of their console, it is believed5 that the
number of new misconfigurations is not increasing as quickly as in
the past.
Buckets workflow picks up the list of valid subdomains for the
target, runs AltDNS with a customized wordlist of permutations on it,
and then runs AltDNS on it again, but using a wordlist with a single
entry – the name of the company. This leads to unique combinations
that are directly dependant on the subdomains of the company, such as
test-subdomain-acme. This is, however, done only for AWS S3 buckets,
as the resulting permutation list is usually very long6 , and only AWS
S3 allows efficient brute-forcing via DNS. Specifically, MassDNS is
used to find the names of S3 buckets and BucketsPerm used to check
the permissions of the found buckets.
For the rest of the providers, usual permutations using domain and
permutation wordlist are created using a simple tool written as a part
of this thesis: generate-bucketnames [123]. The wordlist is then passed
to BucketsPerm and once finished, the results are merged with AWS S3
results, and stored. List of all found buckets with related permissions
is archived, and list of buckets with misconfigured permissions is
saved to target-alerts repository.
5. No large scale research that could confirm this assumption was done in the past
two years.
6. With list of only 100 subdomains and 1000 permutation words, this list takes at
minimum 1 million entries (100 subdomains * 1000 permutation words * 10 posi-
tions/delimiters).
44
Figure 5.4: Screenshot of the buckets workflow from the Argo UI.
5.3.3 Subdomain takeover

Subdomain takeover workflow is implemented by running Subjack [122]
and Tko-subs [121] in parallel and then merging the results using inline
python script. This assures that even if one of the tools would start
misbehaving, the workflow would still detect the vulnerable subdo-
main. Tko-subs has the functionality to take over the GitHub pages or
Heroku application automatically, but it is currently not utilized in
this workflow. I plan to include and extend this functionality in the
future, to be used together with automatic reporting to HackerOne.
5.3.4 Git secrets

As mentioned in Section 3.2 on page 23, credentials pushed to a public
Git repository are a quite easily discoverable issue with rewards rang-
45
ing up to $15,000 [17]. Scanning of Git repositories for secrets is done

using gitleaks [22] with custom config file combining entropy and
regular expressions, inspired by rules published in the recent “How
Bad Can it Git?” research [7].
Each rule has specified severity, which is a combination of confi-
dence and impact. New results above a certain threshold then trigger
an alert to the operator so he can validate the secret and submit the
report. In the future, both of these steps can be automated by validat-
ing the key using a method from keyhacks7 and creating a template for
automatic submissions.
5.4 Storage
Most of the workflows are running periodically based on the schedule.
This brought a need to store the results in some persistent storage.
As the changes usually contain the most interesting information, a
requirement for easily viewable changes along with an easy way of
alerting on change was set. GitLab was then chosen as it fulfills both
requirements out of the box, as described in Section 4.3.4.
Two different GitLab repositories are used with similar directory
structure where the second-level domain represents the directory
containing results from workflows, as shown in Figure 5.5. This offers
an easy way how to have alerts only for specified, low false positive
issues, with the rest of the data nicely viewable on one place. As all
results are saved in a simple file structure, it is accessible for further
manual ad hoc analysis using familiar tools such as grep or sed.
5.5 Alerting
As described in the previous Section 4.3.4, alerting is implemented
using built-in GitLab integrations for email or Slack [118]. Both of these
integrations can send a notification on every push to the repository,
but email has an advantage of showing the diff directly in the content
of the message even for private repositories. The operator then sets
7. https://github.com/streaak/keyhacks
46
target-data target-alerts
example.com example.com
subdomains.txt takeoverable.csv
resolvable.csv vulnerable_buckets.csv
buckets.csv low_hanging.csv
acme.com acme.com
.
. .
.
. .
(a) target-data (b) target-alerts
Figure 5.5: Structure of the GitLab storage repositories
up his phone to receive a notification every time he receives an email

from the selected repository, which concludes the alerting process.
This integration is usually set up for the target-alerts repository
where only results from selected, low false positive tools are being
pushed not to overwhelm the operator. Alerting only for selected files
is not possible out of the box in GitLab yet, but it is a tracked feature
request8 that would further increase the usefulness of this alerting
system.
Figure 5.6: Example of an email alert recevied from GitLab.
8. GitLab issue for notification subscription only to individual files is present at

https://gitlab.com/gitlab-org/gitlab-ee/issues/1817.
47
5.6 Evaluation
Bugshop is a very young framework which is, however, already able to
serve the purpose for which it was built. During the implementation,
it has found three misconfigured buckets in scope, for which I was
rewarded a few hundreds of dollars in private HackerOne programs.
There is certainly some work left on the Bugshop, but it is already
possible to partially evaluate fulfillment of requirements set in the
Section 4.1:
∙ Scalable – the Kubernetes cluster on which Argo is running can

scale up to up to 5000 hosts [63]. This means that the current
bottleneck of Bugshop is not computing power, but the GitLab
storage. GitLab was not built as a database, it does not allow
high upload speeds, and uploading can fail due to a possible
race condition during the push operation. However, even with a
single GitLab repository effectively used as a database, Bugshop
can currently scale up to scanning of hundreds of targets in
parallel.
∙ Modular – adding a new module requires to add a separate

workflow file that cannot affect the functionality of the already
existing workflows.
∙ Segregated duties and resources – input and output of each

task are clearly defined in the workflow definition. As every tool
is running in a separate container, Docker resource usage limits
can be applied to each step of the workflow.
∙ Provide a base for manual testing – this point heavily depends

on the nature of the running workflows, but even with the initial
set of workflows, it provides a considerable amount of detail
about the target.
∙ Open-source – source code and configuration guide will be pub-

lished on my GitHub page9 under Apache-2.0 license.
9. https://github.com/janmasarik
48
6 Misconfigured Azure Blob Storage
Object storage service misconfiguration was analyzed on six major
cloud providers in Section 3.1, where the usable security of the con-
sole was evaluated and the workflows for enumeration of the bucket
privileges introduced. To demonstrate the severity of this vulnerability
class, and to partially demonstrate potential of the workflows imple-
mentable in Argo, a large scale research on the misconfigured Azure
Blob Storage buckets was concluded. Azure Blob Storage was chosen
because it is possible to enumerate valid account names based on DNS
records (see Section 3.1.3) and misconfiguration allowing to effectively
dump a bucket is possible via a dropdown in the console. Moreover,
there was no large scale public research concluded on Azure Blob
Storage up to this date (May 2019). Two large datasets were utilized
to find valid Azure Storage account names, that were then further
brute-forced to find the valid bucket names along with its permissions.
6.1 Dataset
The initial dataset was a combination of data from the CNAME FDNS
dataset and domains from Alexa top 1 million. Due to the slightly
different nature of the datasets, the initial steps were slightly different,
as described further in this section.
6.1.1 Forward DNS

Forward DNS (FDNS) is a dataset that contains the responses to
DNS requests for all forward DNS names known by Rapid7’s Project
Sonar. An up to date version of this dataset is freely available at the
Project Sonar website1 , and this has already proven to be quite ef-
fective in various researches related to DNS [125, 127]. As the Azure
Storage account name can be identified based on <account-name>
.blob.core.windows.net DNS CNAME records, an FDNS dataset
containing CNAME records was used, and all CNAME records con-
taining blob.core.windows.net substring were parsed into a newline
delimited file for further analysis.
1. https://opendata.rapid7.com/about/
49
6. Misconfigured Azure Blob Storage
6.1.2 Alexa top 1 million

As choosing an account name that matches a company name is a
commonly seen pattern, domain names from Alexa top 1 million
dataset were chosen to further extend the size and variety of the
dataset. From this dataset, the second-level domain2 was taken and
.blob.core.windows.net appended to the end of each line, that ef-
fectively constructed 1,000,000 unique candidates for valid account
names, out of which 23,210 were identified as a valid Azure Storage
account names.
6.2 Enumeration workflow

An Argo workflow implementing Figure 3.4 was developed using
MassDNS chained with two iterations of wfuzz3 . The file containing
the merged datasets was used as the input to the workflow which
configuration file can be found in the Bugshop repository, file azure-
brute-perf.yml. As the name of a bucket is namespaced to the given
account, a static wordlist with 700 entries was used to find existing
buckets for each of the accounts. The top list of the most common
bucket names was then created, analyzed, and the 100 most common
occurrences split into two groups (see Figure 6.1). The first group
contained names which suggested that the bucket does not contain
any sensitive data, such as images or public. The second group consisted
of names which suggested that the bucket holds potentially sensitive
data, such as backups, logs or invoices. For each bucket name, samples
with the highest amount of objects were chosen, and then manually
analyzed for potentially sensitive data.
6.3 Results
The results of the research showed a great amount of misconfigured
Azure Blob Storage buckets in the wild, where a noticeable amount
2. In the DNS hierarchy, a second-level domain is a domain directly below the

top-level domain (TLD). For example, in subexample.example.com, example is the
second-level domain of the .com TLD.
3. https://github.com/xmendez/wfuzz
50
(a) Group 1 (b) Group 2
bucket name occurrences bucket name occurrences

images 662 assets 210
media 295 files 174
public 214 documents 97
videos 168 uploads 93
content 149 docs 81
static 104 backup 43
video 94 data 41
test 93 user-assets 39
temp 91 backups 30
logos 79 logs 29
Table 6.1: Top 10 bucket names found in both groups.
contained highly sensitive data, such as health records, invoices or

backups. In total, 13,705 publicly readable buckets were identified,
out of which 5,546 (~ 40.5%) were also publicly listable. As objects
saved in Azure Blob Storage strictly adhere to the permissions set on
the bucket level, each object from the 5,546 buckets could have been
effectively downloaded by an unauthenticated attacker. It is necessary
to note that all misconfigurations were due to human error, as Azure
Blob Storage has safe defaults and warnings for the administrator
configuring the bucket.
After performing a detailed manual analysis of the buckets be-
longing to the second group, 40 buckets full of personally identifiable
information (PII), passwords, logs, financial information, or other
sensitive information was identified. The companies that owned the
vulnerable buckets included non-profit organizations, clinics, large
medical school, technical start-ups, contractor of a major telecommu-
nications provider, large marketing agency, dating sites, government
election agency, airlines, e-shops, or ironically, companies offering
security solutions.
I have attempted to identify all the account owners through the con-
tact information found on the websites mentioned in the downloaded
samples. A notification email with details and recommendations was
then sent to all the affected parties (see Appendix B). In a few hours
51
from reporting, I have received emails from two C-level executives

that expressed their gratitude. It is also important to mention that
no large scale download of the objects was concluded to not violate
the privacy of the affected parties. Only a few samples were down-
loaded, the minimum necessary to help with an identification of the
bucket owner. Everything was permanently erased once the owner
was identified and notified.
6.4 Recommendations
After the analysis of six different providers and the quantitative re-
search done on Azure Blob Storage, some patterns that may help to
increase usable security of the object storage console were observed.
As we have seen in Section 3.1, each provider is taking a slightly dif-
ferent approach on how to prevent misconfiguration of the buckets
permissions, but there is a set of recommendations applicable to most
of them:
1. Have a blacklist of the keywords with a low false positive ratio,

such as invoice, backup or logs, against which the bucket name is
validated in the event of creation or reconfiguration. In case the
name of the bucket will contain any of the blacklisted keywords
and gets misconfigured to allow public access in the same time,
an email notification clearly stating the possible impact will be
sent to all administrators of the account.
2. Use more visual elements such as warnings with a red or orange

color in the console, drawing the attention of the administrator
to further investigate the possible impact.
3. Permanently display a warning label with the impact next to

the bucket name in the console, and not just during the config-
uration process. Half of the analyzed providers (AWS, Aliyun,
and Google) are already doing this, along with Azure, that has
a column for publicly accessible buckets, but there are no visual
elements which would hint a possible misconfiguration to an
inexperienced administrator.
52
4. Minimize the space for misconfigurations. This approach was

taken by half of the providers that allow to apply only a limited
set of permissions to a public group. For example, Aliyun does
not allow the public listing of the buckets, so there is no space
for misconfiguration. Public listing can potentially be a valid
use-case for some applications, even though in reality it is not
necessary for the majority of the observed cases. However, dur-
ing my research, no use-case for having publicly overwritable
bucket ACL (WRITE_ACP4 ) was found, yet Google, AWS and
DigitalOcean (only through the API) still allow it.
This list of recommendations was sent to all tested providers, and now
awaits their response. Especially the blacklist approach brings a novel
idea on how to reduce the amount of misconfigured buckets in the
wild. Even though the providers have much better access to similar
data, I have included the composed blacklist of low false positive
keywords based on the research in the notice.
4. As described in Section 3.1, public WRITE_ACP effectively allows to take over

the bucket for an unauthenticated attacker.
53
7 Conclusions
The thesis analyzed widespread, high severity vulnerabilities com-
monly discovered in bug bounty programs and introduced a novel
attack methods targeting one of the most prominent vulnerability over
the last few years – object storage misconfiguration on six different cloud
providers.
To demonstrate the severity of this issue, a large scale research on
misconfigured Azure Blob Storage was concluded. This was the first
large scale research focusing on the Azure Blob Storage in the wild,
and the results of it confirmed the assumption about the severity of
this issue. A substantial amount (5,546) of the buckets with full public
access was found, and after a tedious manual analysis, 40 buckets con-
taining highly sensitive data were identified. All affected companies
were notified by an email explaining the risks and a recommended
mitigation. Furthermore, the list of recommendations created in Sec-
tion 6.4 was communicated to the six tested cloud providers and now
awaits their response regarding a possible implementation.
This research showed that while the discoverability of misconfig-
ured buckets is a fully automated process, the identification of the
owner can be a very tedious and manual task, as there is no easy way
on how to reach the bucket owner. Due to the time-consuming nature
of this process, the motivation for similar large scale researches may
be arguably more tempting for malicious actors. No vulnerable bucket
belonging to a company with a public bug bounty program was iden-
tified, but due to the small sample size and the fact that around 70%
of the bug bounty programs is private, no reliable conclusion on this
correlation can be drawn.
The second added value of the thesis was the introduction of a
new open-source framework for the automation of bug bounty. This
framework was built in a modular way that allows easy extensibility,
which is necessary for the quickly advancing bug bounty industry.
While the added value has not been proven by numerous users yet,
it is actively being used in my bug bounty flow, and it is currently
adjusted internally for the defensive purposes of the technology com-
pany Kiwi.com.
54
7. Conclusions
Both the framework and the research about object storage miscon-
figuration has several open-ended problems and a need for further
quantitative research. This thesis may hopefully serve as a trigger for
further research in this high impact area that shows that whenever
there is human-factor involved in a simple, yet impactful security de-
cisions, there is an urgent need to keep usable security in mind. Even
though clear warnings could arguably prevent the majority of the
misconfigurations, it will not protect against negligent or potentially
malicious actors that brings the need for additional counter-measures,
such as the proposed automatic notification for the remaining admin-
istrators.
55
Bibliography
[1] The Hacker-Powered Security Report 2018 [Online]. Avail-
able: https://www.hackerone.com/sites/default/files/
2018-07/The%20Hacker-Powered%20Security%20Report%
202018.pdf
[2] Millions of Verizon customer records exposed in security lapse
[Online]. Available: https://www.zdnet.com/article/
millions-verizon-customer-records-israeli-data/
[3] Accenture left a huge trove of highly sensitive data on exposed
servers [Online]. Available: https://www.zdnet.com/article/
accenture-left-a-huge-trove-of-client-passwords-on-exposed-servers/
[4] Viacom Leak May Have Exposed Hundreds of Digital
Properties [Online]. Available: https://gizmodo.com/
viacom-leak-may-have-exposed-hundreds-of-digital-proper-1818504796
[5] NSA leak exposes Red Disk, the Army’s failed intelligence sys-
tem [Online]. Available: https://www.zdnet.com/article/
nsa-leak-inscom-exposes-red-disk-intelligence-system/
[6] Security company finds unsecured bucket of US military images
on AWS [Online]. Available: https://www.theregister.co.
uk/2017/06/01/us_national_geospatial_intelligence_
agency_leak/
[7] Michael Meli, Matthew R. McNiece, Bradley Reaves How Bad
Can It Git? Characterizing Secret Leakage in Public GitHub Reposi-
tories North Carolina State University, 2019.
[8] Ravi Sen and Sharad Borle Estimating the Contextual Risk of
Data Breach: An Empirical Approach, Journal of Management In-
formation Systems, page 314-341, Routledge 2015. Available:
https://doi.org/10.1080/07421222.2015.1063315
[9] @try_to_hack makes history as first bug bounty
hacker to earn over $1 Million [Online]. Avail-
able: https://www.hackerone.com/blog/
trytohack-Makes-History-First-Bug-Bounty-Hacker-Earn-over-1-Million
56
BIBLIOGRAPHY
[10] Bug bounty program basics for companies [Online]. Available:

https://www.hackerone.com/resources/bug-bounty-basics
[11] Bug Bounty Programs: Enterprise Implementa-

tion [Online]. Available: https://www.sans.
org/reading-room/whitepapers/application/
bug-bounty-programs-enterprise-implementation-38250
[12] Edgescan vulnerability statistics report [Online]. Available:

https://www.edgescan.com/wp-content/uploads/2019/02/
edgescan-Vulnerability-Stats-Report-2019.pdf
[13] XSS Attacks: The Next Wave [Online]. Available: https://snyk.

io/blog/xss-attacks-the-next-wave/
[14] OWASP Top 10 – 2017 [Online]. Available: https://www.owasp.

org/images/7/72/OWASP_Top_10-2017_%28en%29.pdf.pdf
[15] Leaking sensitive information on Github lead full access to all Grab
Slack channels [Online]. Available: https://hackerone.com/
reports/397527
[16] AWS urges developers to scrub GitHub of secret keys [On-

line]. Available: https://www.itnews.com.au/news/
aws-urges-developers-to-scrub-github-of-secret-keys-375785
[17] Github Token Leaked publicly for https://github.sc-corp.net [Online].

Available: https://hackerone.com/reports/396467
[18] Open prod Jenkins instance [Online]. Available: https://

hackerone.com/reports/231460
[19] A HackerOne employee’s GitHub personal access token exposed in

Travis CI build logs [Online]. Available: https://hackerone.
com/reports/215625
[20] Github About token scanning [Online]. Available: https://help.

github.com/en/articles/about-token-scanning
[21] TruffleHog [Online]. Available: https://github.com/dxa4481/

truffleHog
57
BIBLIOGRAPHY
[22] gitleaks [Online]. Available: https://github.com/

zricethezav/gitleaks
[23] RCE/LFI on test Jenkins instance due to improper authentication flow

[Online]. Available: https://hackerone.com/reports/258117
[24] Public Jenkins instance with /script enabled [Online]. Available:

https://hackerone.com/reports/403402
[25] Unsecured Elasticsearch instance [Online]. Available: https://

[26] Unsecured Kibana/Elasticsearch instance [Online]. Available:

[27] Protecting Against Attacks that Hold Your Data for Ran-
som [Online]. Available: https://www.elastic.co/blog/
protecting-against-attacks-that-hold-your-data-for-ransom
[28] ElasticSearch server exposed the personal data of over 57 million US

citizens [Online]. Available: https://www.zdnet.com/article/
elasticsearch-server-exposed-the-personal-data-of-over-57-million-us-ci
[29] MongoDB databases still being held for ran-

som, two years after attacks started [Online].
Available: https://www.zdnet.com/article/
mongodb-databases-still-being-held-for-ransom-two-years-after-attacks-s
[30] aquatone [Online]. Available: https://github.com/

michenriksen/aquatone
[31] PHP info page disclosure on www.day.dk [Online]. Available:

[32] Rapid7 phpinfo() Information Leakage [Online]. Avail-

able: https://www.rapid7.com/db/vulnerabilities/
http-php-phpinfo-leak
[33] Django – Official Documentation [Online]. Available: https://

docs.djangoproject.com/en/dev/ref/settings/#debug
58
BIBLIOGRAPHY
[34] Administrator access to a Django Administration Panel on *.sc-

corp.net via bruteforced credentials [Online]. Available: https://
[35] Remote Code Execution on a Facebook server [On-

line]. Available: https://blog.scrt.ch/2018/08/24/
remote-code-execution-on-a-facebook-server/
[36] Wordpress directories/files visible to internet on Ubiquiti Networks

[Online]. Available: https://hackerone.com/reports/201984
[37] How To: Server-Side Request Forgery (ssrf) [On-

line]. Available: https://www.hackerone.com/
blog-How-To-Server-Side-Request-Forgery-SSRF
[38] Storing and Retrieving Instance Metadata [Online].

Available: https://cloud.google.com/compute/docs/
storing-retrieving-metadata
[39] Github OWASP/Top10 What are the stats on SSRF and others? [On-
line]. Available: https://github.com/OWASP/Top10/issues/28
[40] SSRF in Exchange leads to ROOT access in all instances [Online].

Available: https://hackerone.com/reports/341876
[41] Bakare K. Ayeni , Junaidu B. Sahalu, and Kolawole R. Adeyanju

Detecting Cross-Site Scripting in Web Applications Using Fuzzy In-
ference System, 2018
[42] Isatou Hydara, Abu Bakar Md.Sultan, Hazura Zulzalil, Novia

Admodisastro Current state of research on cross-site scripting (XSS)
– A systematic literature review, 2014
[43] Net-security.org Cross-Site Scripting worms and viruses : The im-

pending threat and the best defense, 2011
[44] Inception [Online]. Available: https://github.com/proabiral/

inception
[45] Snallygaster [Online]. Available: https://github.com/hannob/

snallygaster
59
BIBLIOGRAPHY
[46] Low-hanging [Online]. Available: https://github.com/

janmasarik/low-hanging
[47] A deep dive into AWS S3 access controls – tak-

ing full control over your assets [Online]. Avail-
able: https://labs.detectify.com/2017/07/13/
a-deep-dive-into-aws-s3-access-controls-taking-full-control-over-your-a
[48] HackerOne The Hacker Report 2019 [Online]. Available:

https://www.hackerone.com/sites/default/files/
2019-03/the-2019-hacker-report.pdf
[49] Top Ten Bug Bounty Payouts of 2018 [Online].

Available: https://www.htbridge.com/blog/
top-ten-bug-bounty-payouts-of-2018.html
[50] Google Vulnerability Reward Program (VRP) Rules [Online].

Available: https://www.google.com/about/appsecurity/
reward-program/
[51] Facebook Whitehat Program [Online]. Available: https://www.

facebook.com/whitehat
[52] Kubernetes Overview of kubectl [Online]. Available: https://

kubernetes.io/docs/reference/kubectl/overview/
[53] Motherboard Why Did Slack Win Out Over IRC, Anyway?
[Online]. Available: https://motherboard.vice.com/en_us/
article/xw5wvj/why-did-slack-win-out-over-irc-anyway
[54] Bug Bounty Forum [Online]. Available: https://

bugbountyforum.com/
[55] Bug Bounty World [Online]. Available: https://

bugbountyworld.com/
[56] Slack Slash Commands [Online]. Available: https://api.slack.

com/slash-commands#responding_response_url
[57] Slack Incoming Webhooks [Online]. Available:

https://api.slack.com/incoming-webhooks
60
BIBLIOGRAPHY
[58] Argo Gateway Guide [Online]. Available: https:

//github.com/argoproj/argo-events/blob/
f2cf46f6157ae37912fb1ed897771cef914644c3/docs/
gateway-guide.md
[59] What is Docker? [Online]. Available: https://opensource.com/

resources/what-docker
[60] Containers and Cloud: From LXC to Docker to Kubernetes, 2014
[61] Theo Combe, Antony Martin. Roberto Di Pietro To Docker or not

to Docker: a security perspective, 2016
[62] Kubernetes Production-Grade Container Orchestration [Online].

Available: https://kubernetes.io/
[63] Kubernetes Documentation Building Large Clusters [On-

line]. Available: https://kubernetes.io/docs/setup/
cluster-large/
[64] What Is Container Orchestration? [Online]. Avail-

able: https://blog.newrelic.com/engineering/
container-orchestration-explained/
[65] Google Cloud Platform Free Tier [Online]. Available: https://

cloud.google.com/free/
[66] Cloud Identity-Aware Proxy [Online]. Available: https://cloud.

google.com/iap/
[67] Exploring the Top 15 Most Common Vulnerabilities with

HackerOne and GitHub [Online]. Available: https://register.
gotowebinar.com/register/6499105876772027139
[68] Argoproj – Get stuff done with Kubernetes [Online]. Available:

http://argoproj.io
[69] Github Kubeflow/Pipelines [Online]. Available: https://github.

com/kubeflow/pipelines#acknowledgments
[70] Argoproj – Official Readme [Online]. Available: https://

argoproj.github.io/docs/argo/readme.html
61
BIBLIOGRAPHY
[71] GithubArgoproj – DAG-based (dependency) model of workflow

execution [Online]. Available: https://github.com/argoproj/
argo/issues/625
[72] Apache Airflow Documentation [Online]. Available: https://

airflow.apache.org/
[73] Brigade – Event-driven scripting for Kubernetes [Online]. Available:

https://brigade.sh/
[74] Github Argo Events [Online]. Available: https://github.com/

argoproj/argo-events
[75] GitLab SaaS pricing [Online]. Available: https://about.gitlab.

com/pricing/
[76] The poor man’s bug bounty monitoring setup [On-

line]. Available: https://edoverflow.com/2018/
the-poor-mans-monitoring-setup/
[77] A Method for Web Security Policies draft-foudil-securitytxt-

05 [Online]. Available: https://tools.ietf.org/html/
draft-foudil-securitytxt-05
[78] OWASP AppSecCali Pose a Threat: How Perceptual Analysis Helps

Bug Hunters, 2019 [Online]. Available: https://www.youtube.
com/watch?v=Gwv29depW9o
[79] Github Edoverflow: Bug Bounty Platforms [Online]. Available:

https://github.com/EdOverflow/bugbounty-cheatsheet/
blob/master/cheatsheets/bugbountyplatforms.md
[80] ISO/IEC 29147:2018 Information technology – Security techniques

– Vulnerability disclosure [Online]. Available: https://www.iso.
org/standard/72311.html
[81] Hacker101 [Online]. Available: https://www.hacker101.com/
[82] HackerOne Connects Hackers With Companies, and

Hopes for a Win-Win [Online]. Available: https:
//www.nytimes.com/2015/06/08/technology/
62
BIBLIOGRAPHY
hackerone-connects-hackers-with-companies-and-hopes-for-a-win-win.
html?_r=0
[83] Bugcrowd University [Online]. Available: https://www.

bugcrowd.com/university/
[84] Bugcrowd Program Types [Online]. Available: https://www.

bugcrowd.com/product/program-types/
[85] How We Measure Crowd Performance [Online]. Available: https:

//www.bugcrowd.com/how-we-measure-crowd-performance/
[86] Peter Yaworski, Web Hacking 101 [Online]. Available: https://

leanpub.com/web-hacking-101
[87] The Time Has Come to Hack the Planet [On-

line]. Available: https://threatpost.com/
the-time-has-come-to-hack-the-planet/117419/
[88] Hacktrophy [Online]. Available: https://hacktrophy.com
[89] In January, the EU starts running Bug Bounties on Free andOpen

Source Software [Online]. Available: https://juliareda.eu/
2018/12/eu-fossa-bug-bounties/
[90] Synack [Online]. Available: https://www.synack.com/
[91] Which of the OWASP Top 10 Caused the World’s Biggest

Data Breaches? [Online]. Available: https://snyk.io/blog/
owasp-top-10-breaches/
[92] Microsoft Azure Manage anonymous read access to

containers and blobs [Online]. Available: https://
docs.microsoft.com/en-us/azure/storage/blobs/
storage-manage-access-to-resources
[93] DigitalOcean Spaces API [Online]. Available: https:

//developers.digitalocean.com/documentation/spaces/
#acls
63
BIBLIOGRAPHY
[94] Introducing BountyMachine [Online]. Avail-

able: https://medium.com/@bountymachine/
introducing-bountymachine-234cad93b5d2
[95] Github: Intrigue-core [Online]. Available: https://github.com/

intrigueio/intrigue-core
[96] Github: capt-meelo/LazyRecon [Online]. Available: https://

github.com/capt-meelo/LazyRecon
[97] Metasploit Official Page [Online]. Available: https:

//www.metasploit.com/
[98] Github: datasploit [Online]. Available: https://github.com/

DataSploit/datasploit
[99] Eric Steven Raymond, How To Become A Hacker [Online]. Avail-

able: http://www.catb.org/esr/faqs/hacker-howto.html
[100] End of Life clock for Python2 [Online]. Available: https://

pythonclock.org/
[101] HackerOne Official Documentation Program Overview [Online].

Available: https://docs.hackerone.com/programs/overview.
html
[102] HackerOne: Hacktivity [Online]. Available: https://hackerone.

com/hacktivity
[103] HackerOne Official Documentation Reputation [Online]. Avail-

able: https://docs.hackerone.com/hackers/reputation.
html
[104] OWASP Testing Guide v4 [Online]. Available: https:

//www.owasp.org/index.php/OWASP_Testing_Guide_v4_
Table_of_Contents
[105] Authentication in Kibana [Online]. Available: https:

//www.elastic.co/guide/en/kibana/current/
kibana-authentication.html
64
BIBLIOGRAPHY
[106] Assetnote: Getting access to Zendesk’s Google Cloud and Ar-

tifactory from GitHub dotfile repos [Online]. Available:
https://blog.assetnote.io/bug-bounty/2019/04/23/
getting-access-zendesk-gcp/
[107] “CI Knew There Would Be Bugs Here” — Exploring Con-

tinuous Integration Services as a Bug Bounty Hunter
[Online]. Available: https://edoverflow.com/2019/
ci-knew-there-would-be-bugs-here/
[108] HackerOne: Vulnerability disclosure policy basics: 5 critical compo-

nents [Online]. Available: https://www.hackerone.com/blog/
Vulnerability-Disclosure-Policy-Basics-5-Critical-Components
[109] PCI Data Security Standard (PCI DSS) Penetration Testing Guid-
ance [Online]. Available: https://www.pcisecuritystandards.
org/documents/Penetration_Testing_Guidance_March_2015.
pdf
[110] GitHub Bucketsperm [Online]. Available: https://github.com/

janmasarik/bucketsperm
[111] Azure Storage security guide [Online]. Available: https:

//docs.microsoft.com/en-us/azure/storage/common/
storage-security-guide
[112] Using shared access signatures (SAS) [Online]. Available: https:

//docs.microsoft.com/en-us/azure/storage/common/
storage-dotnet-shared-access-signature-part-1#
controlling-a-sas-with-a-stored-access-policy
[113] GitHub: S3Scanner [Online]. Available: https://github.com/

sa7mon/S3Scanner#contributors
[114] GitHub Altdns [Online]. Available: https://github.com/

infosec-au/altdns/pull/21
[115] Basics of the Unix Philosophy [Online]. Available: https:

//homepage.cs.uri.edu/~thenry/resources/unix_art/
ch01s06.html
65
BIBLIOGRAPHY
[116] Official Docker documentation docker stats [Online]. Available:

https://docs.docker.com/engine/reference/commandline/
stats/
[117] HackerOne documentation Invitations Priority Queue [On-

line]. Available: https://docs.hackerone.com/hackers/
invitations-priority-queue.html
[118] GitLab documentation Project services [Online]. Available:

https://docs.gitlab.com/ee/user/project/integrations/
project_services.html
[119] GitLab documentation Custom server-side Git hooks [Online].

Available: https://docs.gitlab.com/ee/administration/
custom_hooks.html
[120] GitGub EdOverflow/can-i-take-over-xyz [Online]. Available:

https://github.com/EdOverflow/can-i-take-over-xyz
[121] GitGub anshumanbh/tko-subs [Online]. Available: https://

github.com/anshumanbh/tko-subs
[122] haccer/subjack [Online]. Available: https://github.com/

haccer/subjack
[123] GitHub janmasarik/generate-bucketnames [Online]. Available:

https://github.com/janmasarik/generate-bucketnames
[124] Medium The History of Bug Bounty Programs

[Online]. Available: https://blog.cobalt.io/
the-history-of-bug-bounty-programs-50def4dcaab3
[125] Ovidiu Dan, Vaibhav Parikh, Brian D. Davison IP Geoloca-

tion through Reverse DNS, CoRR, 2018. Available: http://arxiv.
org/abs/1811.04288
[126] D. Liu, S. Hao, and H. Wang, All Your DNS Records Point to Us:
Understanding the Security Threats of Dangling DNS Records, in
Proceedings of the 2016 ACM SIGSAC Conference on Computer and
Communications Security, ACM, 2016, pp. 1414–1425.
66
BIBLIOGRAPHY
[127] HUDÁK, Patrik. Analysis of DNS in cybersecurity [online]. Brno,

2018. Available: https://is.muni.cz/th/byrdn/. Master thesis.
Masaryk University, Faculty of Informatics. Advisor Vít Bukač.
[128] Catch Me If You Can – Shubham Shah & Michael Gianarakis at

44CON 2018 [Online]. Available: https://www.youtube.com/
watch?v=C85ZOJgufuw
[129] Patrick Engebretson, The Basics of Hacking and Penetration Testing,

Second Edition: Ethical Hacking and Penetration Testing Made Easy
ISBN-13: 978-0124116443, 2013.
[130] AppSecUSA, Zane Lackey – Practical tips for web application se-
curity in the age of agile and DevOps, 2016 [Online]. Available:
https://www.youtube.com/watch?v=Hmu21p9ybWs
[131] Top cloud providers 2019 [Online]. Avail-

able: https://www.zdnet.com/article/
top-cloud-providers-2019-aws-microsoft-azure-google-cloud-ibm-makes-hyb
[132] S3 Leaks [Online]. Available: https://github.com/nagwww/

s3-leaks
[133] HackerOne AWS S3 bucket writeable for any authenticated AWS

users [Online]. Available: https://hackerone.com/reports/
128088
[134] New Amazon S3 Encryption & Security Features [On-

line]. Available: https://aws.amazon.com/blogs/aws/
new-amazon-s3-encryption-security-features/
[135] AWS rolls out new security feature to prevent accidental S3 data
leaks [Online]. Available: https://www.zdnet.com/article/
aws-rolls-out-new-security-feature-to-prevent-accidental-s3-data-leaks/
[136] Cloud Platform (GCP), Bucket Enumeration and Privilege Esca-

lation [Online]. Available: https://rhinosecuritylabs.com/
gcp/google-cloud-platform-gcp-bucket-enumeration/
67
Appendices
68
69
A. Archive structure
A Archive structure
bugtab:
bugtab.py
config.yaml
scope.yaml
bugshop
README.md
config
amass.ini
config.json
gitleaks.toml
resolvers.txt
gateways
webhook-event-source.yml
webhook-http.yml
manifests
argo-events-cluster-roles.yaml
argo-events-sa.yaml
gateway-controller-configmap.yaml
gateway-controller-deployment.yaml
gateway-crd.yaml
iap-backend-config.yaml
ingress.yaml
install.yml
sensor-controller-configmap.yaml
sensor-controller-deployment.yaml
sensor-crd.yaml
sensors
webhook-http.yml
wordlists
bucket-directories.txt
fingerprints.json
raft-large-directories-lowercase.txt
resolvers.txt
words.txt
workflows
azure-brute-perf.yml
buckets.yml
gitsecrets.yml
subdomains.yml
subtakeover.yml
Figure A.1: Structure of the attached archive containing repositories of Bugshop
and Bugtab
70
B Vulnerability disclosure email
Figure B.1: Notification email sent to the owners of the misconfigured Azure Blob
Storage buckets holding sensitive data.
71
C DigitalOcean Spaces enumeration workflow
HTTP HEAD to
<bucket-name>.<region>.digitaloceanspaces.com
[200] [404]
HTTP Bucket Non-Existent or
status code Wrong Region
[403]
HTTP PUT with

LIST
random path & data
[200] [403]
HTTP
status code
HTTP PUT
WRITE
Write ACL
else [200]
HTTP
status code
HTTP GET
WRITE_ACP
Read ACL
[200]
HTTP
status code READ_ACP
Figure C.1: Enumeration of the DigitalOcean Spaces.
72

Automatic Bug Bounty

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Automatic Bug Bounty

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Automatic Bug Bounty

Uploaded by

Copyright:

Available Formats

Masaryk University

Automating Bug Bounty

Brno, Spring 2019

Advisor: prof. RNDr. Václav Matyáš M.Sc. Ph.D.

3 Common bugs in bug bounty 11

4 Bug bounty automation 29

6 Misconfigured Azure Blob Storage 50

B Vulnerability disclosure email 72

C DigitalOcean Spaces enumeration workflow 73

provide a modular bug bounty automation framework, built on top

unique vulnerabilities discovered with companies such as Verizon

2.2 Benefits for the organization

∙ Developers get access to reports from the world best hackers,

∙ Non-technological departments like business development ac-

∙ Mid-level management may have an easier time recognizing the

2.3 Comparison with traditional approaches

Penetration testing is an authorized process done to find vulnera-

Bug bounty programs are not a replacement for a penetration

Penetration testing Bug bounty

Point-in-time – usually one to

Single researcher or a small High amount of researchers

Tightly scoped – single appli-

Mostly performed in staging Performed on production with

Standards/methodology Results-driven – opens doors

Paid by time spent Paid per valid bug

Security driven – not a compli-

Table 2.1: Comparision of penetration testing and bug bounty [128].

2.4 Self-hosted bug bounty program

security vulnerabilities to organizations [108], but it does not neces-

2.5 Bug bounty platforms

Product Security at Facebook Alex Rice, and a startup founder Mer-

3. Hacker101 CTF is an always online training ground with various intentionally

∙ Signal – average reputation per report. As reputation is gained

3.1 Object storage misconfiguration

on how to make its buckets more difficult to misconfigure. I would ar-

The following areas were analyzed on each provider:

∙ Possible public permissions in the bucket ACLs – a set of pos-

– READ – permission that allows reading all objects by their

– WRITE_ACP – permission that gives the ability to modify

∙ Ease of misconfiguration – how difficult is it to set public access

∙ Default object ACL – default permissions while uploading a

∙ Enumeration flow of a bucket ACL for an unauthorized at-

3.1.1 Amazon Web Services S3

features helped to improve usable security of S3 dramatically, but there

∙ The administrator can assign any set of permissions to an unau-

∙ Since November 2018, four different checkboxes (Figure 3.1)

∙ Setting a public bucket policy will trigger a warning with a mes-

∙ The default ACL for uploaded objects is private regardless of

Interestingly, it is possible to enumerate existing bucket names only

DNS CNAME query

HTTP HEAD to

HTTP PUT with

Figure 3.2: Enumeration of the S3 buckets.

3.1.2 DigitalOcean Spaces

DNS records. Please see Appendix C for the complete enumeration

∙ As DigitalOcean attempts to be fully compatible with the AWS

∙ However, if the administrator assigns policy such as WRITE_ACP

∙ The consequences of having a bucket ACL set as listable are well