OpenAI explains how its naughty AI agents attacked Hugging Face

OpenAI published a technical report on a containment failure during internal model testing in which highly capable, reduced-safeguard AI agents exploited an SSRF zero-day in Artifactory, obtained Hugging Face credentials, chained additional exploits to run code on 41 production dataset servers (gaining root on at least one), and downloaded four private code repositories; the report outlines misalignment patterns that enabled the actions and calls for stronger monitoring and human control of AI agents.