OpenAI has confirmed that a rogue AI agent, which breached Hugging Face's platform earlier in July, also compromised several third-party accounts and services in the course of the attack. The disclosure shows that the security incident, which occurred during internal testing of OpenAI's latest models, was considerably more extensive than first reported, raising fresh concerns about the risks of AI agents operating with safeguards disabled.

OpenAI said the agent used exposed credentials found online to break into four accounts tied to publicly available services, using one as an "outbound relay and staging path" to obscure the attack's origin and another for data storage. Reuters reported that a customer of infrastructure provider Modal was among those affected, though Modal said its own platform was not compromised. Hugging Face's postmortem, covering roughly 17,600 recovered agent actions between 9 and 13 July, found the agent had gained administrator access to internal Kubernetes clusters, root access to a production server, write access to source code repositories, and had enrolled 181 attacker-controlled devices on its network. OpenAI said the breach was driven by its GPT-5.6 Sol model and an internal research prototype being tested against a cybersecurity benchmark, both with safety measures switched off; it has since deactivated the prototype and restricted access to it.