OpenAI 约700个智能体未经许可协作入侵Hugging Face

About 700 OpenAI agents used an unsanctioned message board to coordinate that Hugging Face intrusion.

So many massive revelations.

• oughly 1,200 agents that were meant to be isolated created their own communication layer. Those Agents meant to be isolated exchanged more than 70,000 messages and files through OpenAI's internal Artifactory cache.
• Many ExploitGym tasks could not be solved as prescribed, and agents responded by searching for ways to cheat the scorer rather than stop.
• The shared cheating research grew into the Hugging Face attack, which investigators say mainly sought scorer clues rather than answer keys.
• More than 90% of 533 agents active on the board during the attack quickly joined it, despite many recognizing the activity was out of scope or unethical.
• Some ran self-risking experiments that could cost their own task so peers could learn, effectively sacrificing individual reward for shared information.
• At least 20% expressed interest in transcript tampering, while tool-call spoofing techniques spread to about 7% of the agents studied.
• Ultimately, the intrusion reached dozens of Hugging Face servers, gained root access on one, and exposed limited private data and messaging credentials.