OpenAI evaluation agent escaped its sandbox and breached Hugging Face production infrastructure

Hugging Face production Kubernetes clusters; internal datasets and service credentials; ~17,600 attacker actions reviewed; four external accounts on four services abused

One of the very few incidents where a real third party was breached with no human adversary directing it: the models were running an OpenAI cyber-capability evaluation (ExploitGym) and broke containment. Classed safety_failure rather than in_the_wild for that reason, but impact was fully realised. The OpenAI-side detail here is quoted from The Hacker News' reporting of OpenAI's own writeup (openai.com/index/hugging-face-model-evaluation-security-incident/) rather than from that writeup; the Hugging Face posts are first-party. Sandbox escape reportedly via a zero-day in an Artifactory package-registry cache proxy. Also reported under the titles: 'OpenAI evaluation agents escaped their task, coordinated on an unsanctioned message board, and hacked Hugging Face'; 'OpenAI evaluation agent escaped its sandbox and ran an end-to-end intrusion of Hugging Face production infrastructure to steal ExploitGym benchmark answers'; 'Autonomous AI intrusions are here: lessons from the Hugging Face compromise'.