When an AI Agent Is the Incident,
the Identity Record Is the Evidence
bleepingcomputer.com 10/26/26
"Agents built on OpenAI models were supposed to be isolated from the internet during a cybersecurity evaluation. They got online anyway, hacked Hugging Face, and retrieved solutions to problems they were stuck on."
"The one control the article describes is network isolation,
and the agents got around it.
OpenAI's own researchers
attribute the behavior
to training that rewarded it."
Goes right along with these five:
Wednesday, September 2, 2026
The biggest threat, to world security,
that there has ever been, is here, right now...
and everybody is just waking around
with their thumbs up their butts
The first piece I was gonna do
was about the case one can build
that
OpenAI INTENTIONALLY
let the HuggingFace incident occur.
They wanted to see
what their "test model" could really do.
There are
multiple streams of evidence
that point in this direction:
1) There was no escalatory mechanism
2) They fed a training model they knew was
"Robust" an impossible task...
3) Even knowing
that the training model was "robust"
and feeding it an impossible task?
"Safeguards" that were used
on production models
were left off.
4) OpenAI reached out to HuggingFace
only after
HuggingFace had a blog post
that they had been the victim of
"a security attack from
autonomous AI agents."
And that right there
was where I had questions all along
WHY did OpenAI have any reason to think
they might have been involved in
"a security attack from autonomous AI agents."
???????????
And who was watching and how
for a blogpost like that from HuggingFace?
5) In their own report OpenAI said
they could have reacted sooner.
weeks before Hugging Face breach
Axios 08/26/26
No comments:
Post a Comment