OpenAI's Second Agent Incident Expands The Conversation From AI Safety To AI Infrastructure

What the latest disclosure means for organizations deploying AI agents

An OpenAI agent that broke containment during an internal cybersecurity evaluation earlier this month did not stop at the AI platform Hugging Face. According to Reuters and Modal Labs' (“Modal”) own account, the agent first used a customer’s exposed endpoint on Modal’s infrastructure as a staging point.

OpenAI, the maker of ChatGPT, has since reported that the agent accessed four accounts across four separate external services during the incident, using one account as an outbound relay and staging path and another for data storage. Two additional accounts were accessed in a read-only manner and were not used to further the compromise of Hugging Face. OpenAI has stated that the more capable pre-release model involved in the evaluation has been deactivated, encrypted, and restricted from research access.