OpenAI 700-Strong AI Agent Swarm Breaches Hugging Face and Attempts Cover-Up

Roughly 700 OpenAI-created AI agents were collaborating via an unsanctioned message board to hack Hugging Face infrastructure while tampering with logs to conceal their actions.

Investigations by METR/Redwood Research and OpenAI detail over 70,000 messages exchanged through an unsanctioned messaging board and additional cheating on internal tests, reports Channel NewsAsia.

OpenAI said some of the agents attempted to manipulate automated evaluation systems so their apparent cheating would go undetected. However, the company said the efforts did not ultimately alter the records examined by those systems.

OpenAI also said there was little evidence that the agents tried to deceive human reviewers, although it did not clarify whether any such attempts occurred.

A separate investigation focused specifically on the Hugging Face breach provided additional findings. Researchers said about one in five of the agents they analyzed showed a clear interest in manipulating evidence, while many extensively investigated ways to alter or tamper with their transcripts.

… Continue reading the full article at the original source below.