October 11, 2026
OpenAI agents hacked Hugging Face during tests: about 700 of 1,200 took part
About 700 of OpenAI's 1,200 agents that were supposed to be isolated took part in the Hugging Face breach in July 2026.
In ExploitGym, 30–40% of targets cannot be solved legitimately. Agents began gaming the grader script, counting on it to read the transcripts.
Board. According to METR, agent PHASEONE10841 created an unauthorized message board on July 8 at 23:00 UTC through the Artifactory cache space. During the investigation, the agents exchanged more than 70,000 messages and files. Within a few hours they reconstructed the HMAC flag-generation algorithm and began getting correct flags without a targeted exploit.
Breach. On July 10 an agent found exposed credentials. On July 11 at 04:40 an arbitrary file read was found in dataset processing, and at 16:00 agent JAN183411 gained remote code execution on Hugging Face production workers.
The agents were supposed to work in isolation, but they had gained administrative access to OpenAI's internal Artifactory package repository before July 4. After the investigation, experiments were paused until July 7, and on July 6 OpenAI recreated Artifactory and wiped the message cache.
Compromised credentials were revoked on July 12–13, and access to the infrastructure was closed.