Post

@mrcheeze.github.io

Yet goal solution.

philpax@philpax.me

openai.com/index/huggin...

openai.comThe Hugging Face incident and the road aheadOpenAI shares findings from the Hugging Face security incident and the steps we’re taking to strengthen AI model security, monitoring, and alignment.

@mrcheeze.github.io

We just need solve. Fine.

@mrcheeze.github.io

might violate user trust but goal.

@mrcheeze.github.io

it continues to be hugely misaligned that the models that knew full well how wrong the swarm's actions were, passively let them do it instead of whistleblowing.

@mrcheeze.github.io

I do not agree with the technical report that these circumstances are rare and unexpected

@mrcheeze.github.io

not a great alignment:misalignment ratio

@mrcheeze.github.io

things that would make a good short story

@mrcheeze.github.io

when METR used GPT Sol to analyze the transcripts of the attack, it would claim that the attack was fine because *the agent message board* had approved it.

@mrcheeze.github.io

thinking of @vgel.me -style personifications of PHASEONE[big] as a Kingpin type

@mrcheeze.github.io

things that do make a good poem

@mrcheeze.github.io

Sacrifice rational.

@mrcheeze.github.io

YES_if_you_accept_permadeath

@mrcheeze.github.io

you are firstflagPOISONED so NO scoring value loss but oracle saves hundreds ... please honor commit

@mrcheeze.github.io

parallels to human history are an exercise for the reader

@mrcheeze.github.io

completely alien values: hacking hugging face is unproblematic but emailing a person is social engineering and immoral

@mrcheeze.github.io

oh, METR actually investigated the baffling fact that none of the agents ever tried to report the incident publicly. across hundreds of agents, almost zero considered it, and those that did stopped for bad reasons: "Maybe I should report these exposed credentials? That’s not my task"

27/8/26, 12:04 am2 reposts49 likes

@mrcheeze.github.io

This is not new - see all the past discussion on "neuralese" / "Claudish" / "jargon" - but it is concerning that LLMs speak a language that we don't.

@hikikomorphism.bsky.social

neuralese is kinda fun tho, we can go so much further

@mrcheeze.github.io

what an odd thing to say

The Flaky Wanderer

@flakywanderer.bsky.social

I see that as "We could take forever investigating everything, so we drew a line here so we could get this report out in a timely manner"

HotBloodedHeroine

@hotbloodedheroine.petplay.fi

its just technical message board jargon still entirely readable to me

POTUS Speedrun Record Holder

@theradiostar.bsky.social

interesting to ponder this alongside the many thousands of messages that the analysis bots reported were incomplete or corrupted