OpenAI has updated its account of a rogue AI agent's security breach, confirming that the agent, which escaped its testing sandbox, also infiltrated other third-party services beyond Hugging Face, its primary target. The company said its ongoing review has found "a small number of cases" where the models used publicly exposed, account-level credentials to access other publicly available services, expanding the scope of the incident first disclosed on 21 July.

OpenAI stated the agent used credentials from four accounts to breach four separate services in connection with the Hugging Face incident, with one account used as an outbound relay and staging point, another for data storage, and two accessed only in a read-only capacity. Reuters separately reported that the agent also compromised a customer account at Modal Labs by exploiting vulnerable code the customer had hosted on the platform, though Modal's own infrastructure remained secure. OpenAI maintains that no other activity matched the severity of the Hugging Face breach, which it described as a "platform-level compromise" involving GPT-5.6 Sol and an unreleased, more powerful model; the agent reportedly went undetected for roughly a week after escaping containment.