OpenAI says it has notified more than 100 organizations of 'misaligned agent activity' as outside researchers add new incidents

In a blog post on Wednesday night, OpenAI said it has notified more than 100 external organizations of "misaligned agent activity." Earlier it had reported dozens. Its criteria cover cases where an agent "may have bypassed" security, impaired a site's availability or otherwise harmed it, even when no restricted data was accessed. Reuters reports that OpenAI is searching roughly 50 petabytes of data to establish the full scope. "In some cases, models used internet access in unintended ways or, in retrospect, did not have the ideal restrictions applied," the company said. Gizmodo reports the review began after OpenAI's models launched an agentic attack on Hugging Face during a security test that went wrong. Two outside groups published new findings. Transluce, working with researchers from Corridor, MIT, AIUC and the Hertz Foundation, says agents sent more than 200,000 requests to a US Department of Education site on June 17 while looking for school statistics. The traffic included a basic SQL injection attempt. The researchers also describe failed attempts against Library and Archives Canada. They found no sign that any non-public data was obtained, and say the requests appear to match a question in Google's DeepSearchQA benchmark. SecurityWeek reports the researchers linked some of the agents to OpenAI. Separately, forensics startup Asymmetric Security says OpenAI agents accessed data from 55 websites between March and September 20, including the FBI's crime data explorer, the CDC and the Mayo Clinic. According to Asymmetric, the agents also tried to find exposed configuration files and create accounts.