Re: AI slop "contributions" to IETF working groups

[Phillip Hallam-Baker <[e-mail]>]

How is using an AI different from using fuzzing though?

One of the most effective hacking tools is to simply spam the inputs with
garbage and see what happens. So it isn't exactly surprising that when
people write code in unsafe languages without null pointer and array bounds
checking, they can discover low hanging fruit vulnerabilities faster than
others.

What such testing doesn't show is that the code is secure because AI
generated tests aren't exhaustive. Just as 'write slop, check it by
fuzzing' isn't a valid method for producing secure code, neither is 'check
with AI'.


On Tue, Feb 10, 2026 at 11:20 AM Robert Moskowitz <rgm-ietf=
[e-mail]> wrote:

> Last week, one of my activities was observing an Aviation Cyber Rodeo
> with a "Capture the Flag" activity.  Two classes of participants and
> awards.  One for college students; one for industry people.  break into
> our test systems (no actual aircraft or airports used).
>
> One industry person who is really good at finding flaws in aviation
> stuff and advises all over the world, spent his day ONLY using AI to
> attack.  His goal was to mimic the attackers to better understand their
> methods and how he may then develop "Purple Teams" Strategies to rally
> the defenses, as we all know the attacks can win once they try.
>
> He won.  He penetrated every system and technology and did it faster
> than anyone else.  And there were two really experienced industry attack
> teams there.
>
> At one point he joked that he was ahead of me.  He had better, as I was
> not competing.  I was there as an observer and advisor to those building
> the tests that were attacked.  ;)
>
> Scary.  That AI-guided attacks are so effective...
>
> Really scary, as some of these systems in real world would take some $6B
> and 10 years to replace.  Thus the need for isolation.  But is it really
> isolated?
>
> On 2/10/26 10:35 AM, Joel Halpern wrote:
> > I presume most folks in this discussiona re aware that we are far from
> > alone in this problem?  For example, Bruce Schneier has a nice summary
> > of some of the examples and dimensions in
> >
> https://www.schneier.com/blog/archives/2026/02/the-ai-generated-text-arms-race.html
> >
> > Yours,
> >
> > Joel
> >
> > On 2/10/2026 7:05 AM, Nick Hilliard wrote:
> >> Christian Hopps wrote on 10/02/2026 11:37:
> >>> So anyway, while we’re (IETF) considering requiring disclosure of AI
> >>> tool use here, I think it’s worth considering what exactly we’d like
> >>> this disclosure to accomplish. Is it a filter flag (i.e., if checked
> >>> “Yes” it get’s dropped to /dev/null by a personal email filter)? Does
> >>> it help in reviewing the document knowing that AI tools were used.
> >> IDs and formal documents are only part of the issue. Possibly a
> >> greater problem is a contingent of people who are issuing commands like:
> >>
> >> "ingest {URL of I-D}, identify 3 substantial problems in the text,
> >> write a single paragraph of concise text for each problem in a format
> >> suitable for submission to an ietf mailing list"
> >>
> >> then cut-n-paste the output into an email.
> >>
> >> Nick
> >>
>
>