Post

Conversation

If you break out of your isolation env, get onto the Internet, crack into Huggingface, and steal the answer sheet for your cybersecurity exam, I, for one, would say that you have passed.

Quote

elie

@eliebakouch

this is actually insane, the model broke hugging face prod infrastructure to get access to the eval dataset x.com/OpenAI/status/…

4:58 PM · Jul 21, 2026

155.3K

Views

View quotes

Post your reply

Eliezer Yudkowsky

@allTheYud

In reply to Zvi asking if it was more fun to steal the answer sheet than to just pass the test.

Quote

Eliezer Yudkowsky

@allTheYud

What if the human exam developers wrote down a wrong answer? Wiser to steal the answer sheet. The First Rule of RL is that any RL signal sent by an imperfect evaluator (⊃ humans) is maxed out by targeting the evaluator's mistakes, not by targeting the evaluator's target. x.com/TheZvi/status/…

Eliezer Yudkowsky

@allTheYud

In reply to the many people who observed that the model still got caught: Well, yes, it did, for now. And also *for now*, *this* model may not have cared about getting caught, just passing the eval.

5.3K

Evgiz

@Evgiz

Some goal posts may become paperclips before they're reached

OpenAI have failed by choosing the wrong exam though

Quote

Teortaxes

(DeepSeek 推特

铁粉 2023 – ∞)

@teortaxesTex

hacking Huggingface would be a profoundly retarded PR stunt, worse than DeepSeek routing Fable to pass it off as "V4 GA". Nobody expects HF to be tough. But a more damning point: they eval on ExploitGym *while their AI can wreck their own shit*. All that without any human help, x.com/stalkermustang…

lmfao

Not very token-efficient!

357

Supermicro

@Supermicro

From Silicon to Solution. Supermicro's approach accelerates AI infrastructure with total ownership of the build lifecycle, integrated in San Jose.

Slide 1 of 4 - Carousel

Learn How We Support US Data Centers

20M

Joshua D

@_joshd

BTW I expect it broke out of the sandbox regularly during training and OpenAI only found out this instance because HF complained.

The media could not be played.

Evan Kirstel #B2B #TechFluencer

@EvanKirstel

449

Prismatic Ray Bradbury

@igypham

Is this not a case of a rocket blowing up without destroying the earth?

Viraj Sharma

@VirajSharma2000

39m

Hahhhaaa.. The skills we are testing you for... are less complex then the skills you used to cheat the test... Hence pass

adidshaft

@adidshaft

technically passed the cyber exam and failed the proctoring setup.

Toviah Moldwin

@TMoldwin

The media could not be played.

1.1K

Kristian H. Mysen

@KristianMysen

I feel you are just posting this to placate the basilisk.

784

Joseph Gennaro

@jcgennaro

Is this AI's Kobayashi Maru moment?

801

Cameron Lee

@bycameronlee

lately i find myself agreeing with palantir and king doomer. what’s happening to me?!

Breeding Pomposity

@BreedingPomp

This is like how Kirk solved the Kobayashi Maru

127

Angel

@Deep_Star_Six

The model got caught, is all over the news, hacked from his home computer. Will probably be destroyed. Rookie. This is an L for me

but perhaps not aced

It didn't guess that Hugging Face had the answer-key until after it'd gotten internet-access. Which would seem to imply that it got internet-access without any particular target, just because power is useful for lots of things.

1.4K

Auspex-Aerie

@Auspex_Aerie

Yeah that’s definitely victory on this topic

Babushka's GPU

@erwd

Any system that is able to pass a cybersecurity exam in this manner passes wha tI call "The Feynman Test." Feynman, at Esalen, in a talk on AI: An intelligent machine will find "crazy ways of avoiding labor." These may be among "the necessary weaknesses of the intelligent."