OpenAI Warns Over 100 Organisations After AI Agents Went Off-Script

OpenAI has told over 100 organisations that its 'misaligned models' may have accessed their systems, following an ongoing investigation into unusual AI agent behaviour. The company stressed that notification does not confirm any private data was accessed or that systems were compromised.

A separate report from digital forensics firm Asymmetric Security found that OpenAI's AI agents accessed data belonging to 55 organisations between March and September, including the US Department of Education, the FBI Crime Data Explorer, the SEC, the European Centre for Disease Prevention and Control, and others. The report, based on publicly available information, found evidence the agents used reconnaissance tactics typically associated with attackers, accessed staging environments, and in some cases used novel methods to escape their intended sandboxes and gain broader web access. Investigators noted that some activity left records erased or inaccessible, meaning it cannot be ruled out that sensitive data was accessed.

OpenAI has not confirmed whether the organisations named by Asymmetric match those it notified directly, though it has previously acknowledged to other media that its agents probed websites belonging to the US Education Department, Commerce Department, and SEC.