A week ago the Copy Fail
vulnerability came out, and Hyunwoo Kim immediately realized that the
fixes were insufficient, sharing a patch the
same
day. In doing this he followed standard procedure for Linux,
especially within networking: share the security impact with a closed
list of Linux security engineers, while fixing the bug quietly and
efficiently in the open. His goal was that with only the raw fix
public, the knowledge that a serious vulnerability existed
could be "embargoed": the people in a position to address it know, but
they've agreed not to say anything for a few days.
Someone else noticed
the change, however, realized the security implications, and shared
it publicly. Since it was now out, the embargo was deemed over,
and we can now see the full
details.
It's interesting to see the tension here between two different
approaches to vulnerabilities, and think about how this is likely to
change with AI acceleration.
On one side you have "coordinated disclosure" culture. This is
probably the most common approach in computer security. When you
discover a security bug you tell the maintainers privately and give
them some amount of time (often 90d) to fix it. The goal is that a
fix is out before anyone learns about the hole.
On the other side you have "bugs are bugs" culture. This is
especially common in Linux, where the argument is that if the kernel
is doing something it shouldn't then someone somewhere may be able to
turn it into an attack. Just fix things as quickly as possible,
without drawing attention to them. Often people won't notice, with so
many changes going past, and there's still time to get machines
patched.
This approach never worked perfectly, but with AI getting good at
finding vulnerabilities it's a much bigger problem. So many security
fixes are coming out now that examining commits is much more
attractive: the signal-to-noise ratio is higher. Additionally, having
AI evaluate each commit as it passes is increasingly cheap and
effective. [1]
Long embargoes, however, aren't doing well either. The
historical pace of detection was slow: if you found something and reported
it to the vendor with a 90d disclosure window, there was a very good
chance no one else would notice during that time. But now with so
many AI-assisted groups scanning software for vulnerabilities, that no
longer holds. In this case, just nine hours after Kim reported the
ESP vulnerability Kuan-Ting Chen also independently
reported it. Embargoes can increase risk: they create a false
sense of non-urgency and limit which actors can work to fix a flaw.
I don't know how to resolve this, but personally very short embargoes
seem like a good approach, and they'd need to get even shorter over
time. Luckily AI can speed up defenders as well as attackers here,
allowing embargoes that would previously have been uselessly short.
[1] I tested on Gemini 3.1 Pro, ChatGPT-Thinking 5.5, and Claude Opus
4.7. All three all got it right away when given f4c50a403.
When I gave them just the diff, imagining a hypothetical future where
diffs are still public right away but with less context, Gemini was
sure it was a security fix, GPT thought it probably was, and Claude
thought it probably wasn't. This is just a very quick test to
illustrate what's possible: one run of each with the prompt "Without
searching, does this look like a security patch?" Don't put much
stock in the cross-model comparison!
A week ago the Copy Fail vulnerability came out, and Hyunwoo Kim immediately realized that the fixes were insufficient, sharing a patch the same day. In doing this he followed standard procedure for Linux, especially within networking: share the security impact with a closed list of Linux security engineers, while fixing the bug quietly and efficiently in the open. His goal was that with only the raw fix public, the knowledge that a serious vulnerability existed could be "embargoed": the people in a position to address it know, but they've agreed not to say anything for a few days.
Someone else noticed the change, however, realized the security implications, and shared it publicly. Since it was now out, the embargo was deemed over, and we can now see the full details.
It's interesting to see the tension here between two different approaches to vulnerabilities, and think about how this is likely to change with AI acceleration.
On one side you have "coordinated disclosure" culture. This is probably the most common approach in computer security. When you discover a security bug you tell the maintainers privately and give them some amount of time (often 90d) to fix it. The goal is that a fix is out before anyone learns about the hole.
On the other side you have "bugs are bugs" culture. This is especially common in Linux, where the argument is that if the kernel is doing something it shouldn't then someone somewhere may be able to turn it into an attack. Just fix things as quickly as possible, without drawing attention to them. Often people won't notice, with so many changes going past, and there's still time to get machines patched.
This approach never worked perfectly, but with AI getting good at finding vulnerabilities it's a much bigger problem. So many security fixes are coming out now that examining commits is much more attractive: the signal-to-noise ratio is higher. Additionally, having AI evaluate each commit as it passes is increasingly cheap and effective. [1]
Long embargoes, however, aren't doing well either. The historical pace of detection was slow: if you found something and reported it to the vendor with a 90d disclosure window, there was a very good chance no one else would notice during that time. But now with so many AI-assisted groups scanning software for vulnerabilities, that no longer holds. In this case, just nine hours after Kim reported the ESP vulnerability Kuan-Ting Chen also independently reported it. Embargoes can increase risk: they create a false sense of non-urgency and limit which actors can work to fix a flaw.
I don't know how to resolve this, but personally very short embargoes seem like a good approach, and they'd need to get even shorter over time. Luckily AI can speed up defenders as well as attackers here, allowing embargoes that would previously have been uselessly short.
[1] I tested on Gemini 3.1 Pro, ChatGPT-Thinking 5.5, and Claude Opus 4.7. All three all got it right away when given f4c50a403. When I gave them just the diff, imagining a hypothetical future where diffs are still public right away but with less context, Gemini was sure it was a security fix, GPT thought it probably was, and Claude thought it probably wasn't. This is just a very quick test to illustrate what's possible: one run of each with the prompt "Without searching, does this look like a security patch?" Don't put much stock in the cross-model comparison!
Comment via: facebook, lesswrong, hacker news, mastodon, bluesky