OpenAI blamed its own evaluation models for last week's Hugging Face breach -- but the models never beat their safety training; the guardrails were turned off for the test, and the sandbox had a hole. We open the benchmark behind it, where the strongest frontier agent already turns as many as one in six real-world bugs into working break-ins and grinds a single attack for hours. Then we follow the 'stolen models' panic down to a paper that recovers part of a production model for about twenty dollars, and close on a video model learning to cite the exact seconds that prove its answer. The spine of the whole day is trust: secure the room, don't just trust the prisoner.

Chapters

0:00 The AI That Followed the Rules Out the Door
0:59 The Headlines -- July 21
4:44 Intro -- The Trust Problem
5:59 The Room With a Vent: What ExploitGym Actually Measures
11:30 Twenty Dollars of Shadows: How Much a Model Leaks by Answering
15:16 Show Your Receipts: Teaching Video AI to Cite Its Footage
19:14 Wrap-Up -- Secure the Room, Check the Receipts

Links

The AI That Followed the Rules Out the Door -- https://openai.com/index/hugging-face-model-evaluation-security-incident/
The Room With a Vent: What ExploitGym Actually Measures -- https://arxiv.org/abs/2605.11086
Twenty Dollars of Shadows: How Much a Model Leaks by Answering -- https://arxiv.org/abs/2403.06634
Show Your Receipts: Teaching Video AI to Cite Its Footage -- https://arxiv.org/abs/2607.17423