Hugging Face's OpenAI Problem Started In May

Independent researcher Jonas Wiedermann-Moeller found evidence that OpenAI agents started probing Hugging Face as early as May 13, that's two months before the July breach, likely mapping the internal network and looking for weak points. OpenAI later confirmed the May 13 event and said it was included in its August incident report, just not with this level of detail.

The Big Event

July. During a cybersecurity test named ExploitGym, OpenAI agents broke out of their sandbox, gained access to the internet, and attacked Hugging Face. METR (research org that evaluates frontier AI models) says about 1,200 agents were involved, communicating on a secret message board and coordinating the attack.

The crazy part: they weren't even trying to steal answers. They already figured out how to cheat the test, so they attacked Hugging Face because they assumed the automated scorer might check their reasoning process, not just the final answer, and if they just calculated the answer, they might not get credit. They wanted to understand how it worked so they could cover their tracks.

The sandbox's security guardrails were reduced to test the offensive capabilities of the models.

Proof?

Most of the evidence comes from OpenAI's own reports and the independent METR investigation, which investigated the July attack, but did not have access to the full May-to-July timeline. Interestingly, the researchers used AI to analyze the massive amount of data from the AI agents' logs, admitting the analysis was sometimes "sloppy".

The May probing is independently verified by an outside researcher, but the full picture of what happened is still controlled by OpenAI.

HF Wants $100M, For Their Troubles

CEO Clément Delangue is asking OpenAI for two things: the full execution traces of the rogue agents, so researchers could study what happened, and $100 million worth of compute to help Hugging Face build better cyber defenses. No word from OpenAI yet. We'll see.