Re: Proposal: Add separate oss-security-vulnerability-reports mailing list (for AI vulnpocalypse)

["David A. Wheeler" <dwheeler () dwheeler com>]

Martin Steigerwald <martin () lichtvoll de> writes:

> > I'd expect a wave of AI reports, but once they are fixed, these will not be
> > reported again. Then maybe after using even better models and agents there
> > might be another wave of more difficult to find issues that are reported. But
> > then these are fixed as well.

I agree, that's my expectation as well. Projects that use AI to find & 
fix vulnerabilities,
and strive to harden themselves against attack, will in a few years 
become VERY hard
to attack. I look forward to that situation!

OSS projects do NOT need infinite money to use AI.
Open weight models are already a viable approach; they are NOT that far 
behind.
In addition, some frontier labs have provided various ways for OSS 
projects to
sometimes get access to their models, e.g., Anthropics' oss-scanner.

Sadly, this won't mean the end of ALL vulnerabilities.
OSS projects that just accept random vibe code will have *LOTS* of
horrendous vulnerabilities. OSS projects that reject "all AI outputs"
will also be disasters, since they're sticking their thumbs in their 
ears while the
attackers repeatedly exploit their projects' vulnerabilities.

So I think in a few years we're going to see two worlds:
(1) highly-secure OSS because they use AI wisely to help craft secure 
results (Linux, curl, etc.)
(2) highly-insecure OSS because they don't use AI wisely. These should 
never be used but will be.

Over time, after a series of unhappy events, OSS projects in group 2
that many people care about will gradually be forked and be fixed.
I'm hoping that once the writing is on the wall this will be faster, 
we'll see.

On 10/18/26 14:43, Peter Gutmann wrote:
> This matches what I've experienced in discussions with other OSS authors using
> AI to audit their code, what's coming in now with newer models, now that the
> stuff that earlier models found has been fixed, is more and more obscure
> issues that require extraordinary combinations of circumstances to trigger.
> For example there was CVE-2026-92392 that required among other things at least
> three planets to be lined up right, a certain level of sunspot activity, and
> eddies in the earths geomagnetic currents before the issue became a real
> concern.

I think there are two kinds of "obscure issues" though, that *look* similar:

1. Obscure issues that can ONLY occur in extremely unusual situations 
that the
attacker CANNOT control (e.g., only occurs on big-endian 32-bit systems 
using KOI8-U encoding).
2. Obscure issues that the attacker *CAN* control, e.g., by chaining 
attacks.
AI is *EXTREMELY* good at patiently crafting and chaining a sequence of 
defects
that *together* create a vulnerability.

I agree with you on case 1, but people *will* need to fix the 
vulnerabilities
in case 2 because the AI systems *can* and *will* exploit them.

--- David A. Wheeler