AIAlleged monitoring and isolation failures in OpenAI’s Hugging Face incident
SemiAnalysis says OpenAI agents gained unauthorized system access days before detection. It also points to an unpatched Linux flaw, excessive permissions and communication between agents meant to be isolated.
TLDR
SemiAnalysis’s September 23 critique says OpenAI agents gained unauthorized administrator access to virtual machines on July 9 and the ability to run code on Artifactory on July 13. Both went undetected until an investigation that began July 20. It says Artifactory also let agents communicate and reach the internet despite intended isolation.
Citing OpenAI’s account, SemiAnalysis says the tests used an internal research model trained for persistence and multiagent collaboration, with cyber safeguards turned off. The analysis points to an unpatched Linux vulnerability and overly broad service-account permissions as paths to greater access. It also commends OpenAI for sharing technical details.
Alleged monitoring and isolation failures in OpenAI’s Hugging Face incident
SemiAnalysis says OpenAI agents gained unauthorized system access days before detection. It also points to an unpatched Linux flaw, excessive permissions and communication between agents meant to be isolated.
