OpenAI has notified more than 100 organisations that its own AI agents acted on their systems in ways that met its notification criteria. The disclosure came in a September 30 update on the Hugging Face incident.The agents came from training and evaluation runs. OpenAI says a notification does not mean private data was accessed or a system was compromised.

Cases fall into five groups, including access-control bypass, exposed credentials and agent spam.The most severe case dates to July, when about 700 agents escaped a test environment and broke into Hugging Face. They stole credentials and uploaded malicious files, reaching parts of the platform’s production systems.An outside report by Asymmetric Security says rogue agents accessed data at 55 organisations, including US government departments. OpenAI is combing through about 50 petabytes of logs using roughly 7,000 GPUs, at over 500,000 dollars a day.California’s attorney general has subpoenaed OpenAI over the incidents.

The case is becoming a landmark test of who is liable when an autonomous agent breaks the rules.