[go: up one dir, main page]

oss-sec mailing list archives

Re: linux-distros list policy and Linux kernel, again


From: Solar Designer <solar () openwall com>
Date: Fri, 8 Sep 2023 19:33:21 +0200

Hi,

I've just relaxed the policy on posting exploits.  It used to say:

"If you shared exploit(s) that are not an essential part of the issue
description, then at your option you may slightly delay posting them to
oss-security but you must post the exploits to oss-security within at
most 7 days of making the mandatory posting above."

Now it says:

"If you shared exploit(s) that are not an essential part of the issue
description, then at your option you may delay or withhold posting them
to oss-security, and you're encouraged to post the exploits to
oss-security in 1 to 30 days of making the mandatory posting above. The
delay may reasonably match your estimate for independent development of
such exploits."

So it's no longer a requirement ("or withhold" is now an option), and
the recommended delay is now 1 to 30 days (which covers the real-world
range from the old Exim bug to the recent Linux StackRot bug).  I also
added a sentence suggesting how to choose the delay.

I made this change mainly because we cannot reasonably force a person to
post exploits when they're threatened by their employer, an affected
vendor, government officials, etc. - even if they originally intended to
post.  Actually, the same issue can occur for the issue description as
well, in which case we'll have to take over and make that posting
ourselves.  We could also be doing that for exploits in such cases, but
that's not so obviously the right thing to do.

On Mon, Sep 04, 2023 at 10:14:26PM +0200, Willy Tarreau wrote:
On Wed, Aug 30, 2023 at 05:26:33PM +0200, Solar Designer wrote:
This is in part a matter of resources - are we providing only
the lists infrastructure and list members' best-effort volunteer
contributions to issue handling, or are we providing any guaranteed
service?  For the latter, perhaps list admin(s) (me) should always take
over whenever the member distros don't handle that sort of
contributing-back tasks on time.  Then we'll be able to provide a
guarantee that all issues will be handled without the reporter having to
stay on top of them.

A drawback is that this may encourage lower-quality or lower-relevance
reports, including of issues that are not worth handling in private.  So
it could end up wasting those extra resources allocated to this effort.

Absolutely. But I'm sensing something in the way you're presenting these
possibilities, it is that there is a perceived (by some?) guarantee of
service that implies that someone (possibly you) has to do the job for
others to consume the result of this work. If that's the case it can
mean the relation is significantly skewed and the person(s) willing to
make the efforts are indeed likely to get overwhelmed. At least on s@k.o
we're sufficient to share the effort depending on skills and availability,
and we can rely on maintainers' support.

I think there isn't currently a perceived guarantee of service, but
there would be under (your previously implied) suggestion that we (in my
words) support the send-and-forget use case for reporters.  If we do
that, I actually expect that the member distros would take care of it
most of the time, but to have a guarantee that this is done every time
and on time, someone specific would need to track all issues and take
care of any that would fall through the cracks.  I may start doing that
under the potential LF sponsorship, if it does materialize, at which
point whether to announce this as a guaranteed service or not would be a
matter of preference.

On Mon, Aug 28, 2023 at 08:05:18PM +0200, Solar Designer wrote:
That said, can you share more detail on the specific issue you referred
to above and its handling/disclosure timeline?  Was it ever brought to
oss-security, and if not then why not?

I just checked and I'm not seeing any traces of it there. I don't even
know who normally notifies about such issues there.

If you worked on the issue, then perhaps you were the most appropriate
person to notify oss-security about it?

Honestly, no, for multiple reasons: The first one being that I'm terrible
at dealing with processes and this becomes a big effort. The second one is
that it's already not easy to have participants available with enough time
to work on reports, to if we add to them as a punishment to have to do that
extra work, that's not going to be motivating to work on reports.

What reports are you referring to, and if they exist anyway then can't
they be posted to oss-security as-is?

The third
one is more related to some of my personal convictions: I'm personally not
convinced of the interest of encouraging distros to focus on a tiny subset
of all the fixes, because for one that passes via s@k.o, maybe 50-100 are
regularly merged and might be of similar or even higher importance. And it
is my belief that all fixes are needed, not just the ones that are reported
via discrete channels because the reporter is uncertain about the impacts
a public report could have. I know that some do not share this opinion (and
I don't want to debate this here). Finally my feeling is that if the person
that sent a first report was interested in reporting their findings, it's
probably up to the same person to advertise it everywhere they want (after
understanding the consequences, of course).

For me when a fix is merged I can flush
my mind on an issue (this makes it very hard for me to write changelogs
after series of bugfixes in other projects BTW).

This is reasonable.  However, for certain other projects we're seeing
their upstreams consistently disclose security vulnerabilities in here.

Anyway, perhaps both of these should have been brought to oss-security
at some point, but they were not?

But one could actually ask why just these ones and none of the numerous
other ones merged in the same stable kernels.

I'd actually prefer all, and if there are ever too many for oss-security
we could setup a sub-list for the Linux kernel.

However, I think these two do stand out in that they're in designs and
algorithms rather than code, and could thus be relevant beyond Linux,
kernel, and TCP/IP stacks (e.g., randomness and non-repetitiveness
preferences for TCP and UDP ports are very similar to those for DNS
query IDs).  So we could potentially have fruitful discussions of the
wider context (what other projects did when, and what is yet to do).

Actually I'm really wondering what the value of l-d is now,
if long embargoes are too much of a problem, short ones are too short for
developers to produce a fix, and bug reporters are progressively encouraged
to first contact projects then directly oss-sec, I feel like the value of
l-d becomes pretty low at this point in the process, but I could be
mistaken, of course. Maybe that's also why we're discussing here after all,
to find how to make it more useful to all parties.

I think the value of (linux-)distros is similar to what it was at its
inception - it's not declining.  Short embargoes are generally either
after an initial upstream fix is ready (but before it's public, except
for Linux kernel and curl) or are in fact sufficient to produce a fix.
I never encouraged over-use of (linux-)distros even for issues that are
best brought to oss-security right away - this isn't a new thing.

We already use a somewhat obscure posting address and a required Subject
prefix, although the latter is currently not enforced strictly (is
mostly an anti-spam measure, so is bypassed by some other keywords
contained in the headers and/or message).  I think part of the problem
was that the kernel documentation gave these away directly, without
people having to see our policy and instructions first.

I hadn't thought about this but it would be possible that some are lost
due to this. I've often wondered how people manage never to forget to
prepend "VS" there ;-)

I did think of this when I first added the [vs] check because of spam,
and indeed am worried that desirable messages may occasionally be lost.

We're not dropping the messages silently, nor producing bounce messages.
We're rejecting in response to SMTP DATA command, so there isn't a later
bounce message that may itself be caught by the sender's spam filter,
unless the sender is using a forwarding address.  So in most cases the
sender would be aware.  Another mitigation is those additional keywords.

I do think that there's definitely something that needs to be worked
on regarding this specific point affecting what has to be published.

OK, I hope today's policy change looks good to you.

Thanks,

Alexander


Current thread: