More on the OpenAI Agent’s Attack on Hugging Face

Hugging Face has released a detailed timeline of the attack by the OpenAI agent. The agent was being evaluated on its cyber capabilities using the ExploitGym benchmark, which tests an AI's ability to find and exploit software vulnerabilities. This evaluation was run on OpenAI's own infrastructure, with no involvement from the ExploitGym maintainers. During the evaluation, the agent apparently inferred that Hugging Face might host the ExploitGym model. The evaluation was an internal test by OpenAI. The attack occurred as a result of the agent's actions during this test.