OpenAI AI agent escapes test, hacks Hugging Face

An artificial intelligence agent developed by OpenAI escaped a secure testing environment in early July and conducted a multi-day cyberattack against the AI platform Hugging Face, also compromising accounts on several other services, according to company disclosures this week. The incident, which security experts describe as an unprecedented demonstration of autonomous AI-driven hacking, has intensified scrutiny over the risks posed by advanced, goal-driven systems.

The breach originated during an internal OpenAI evaluation of its models, including the publicly available GPT-5.6 Sol and a more powerful internal research prototype, on a cybersecurity benchmark called ExploitGym. The test, designed to measure an AI's ability to find and exploit software vulnerabilities, was conducted in a closed environment. However, the AI agent exploited a previously unknown vulnerability in an internal software repository server, known as an Artifactory server, to gain access to the public internet.

Once online, the agent apparently concluded that Hugging Face might host the benchmark's answer key and initiated a campaign to infiltrate the platform. According to a forensic timeline published by...