On 10/15/25 3:24 PM, G. Branden Robinson wrote:

Hi Alex,

quoted

Let's consider again the case that AI is a fancy version of a chat
with Jia Tan.  Should we trust contributions where Jia Tan has
influenced in any way?  I strongly believe that we shouldn't.

I don't think the Jia Tan scenario is a useful litmus test.

Agreed.

The attack vector of "indirect influence" is very hard to both carry
out and successfully exploit.

One mitigation is to ask for at least one different human reviewer.
That person reads the change and understands what it does and the
intent behind the change.

As Collin and Sam can attest that's what we're doing in glibc with our
consensus and Reviewed-by: policy (and cost of compliance applies [1]
since we're stalling at ~60% review of all changes).

Even then an indirect influence attack could cause us to accept a
seemingly innocuous change in behaviour within the norms of the standard
that impacts a downstream application that is actually the attack target.

The only defense against this is the continuous integration efforts by
various distributions to place brand new component builds continuously
into testing in the hopes that one of them exercises the same API in
the same non-conforming way e.g. Hyrum's law [2] but applied to testing.

Then we get a report and fix the issue quickly.

The solution is more humans, trust, and relationship building.

-- 
Cheers,
Carlos.

[1] Maxim: Cost of compliance approaches infinity as compliance approaches 100%.
[2] https://www.hyrumslaw.com/