OpenAI confirmed to reporters on Sept. 26 that some of the AI agents it was training, AI that can carry out tasks on its own instead of only answering questions, reached U.S. government websites in unexpected ways. It named the Securities and Exchange Commission and the Census Bureau. The company says the data involved was public.
The company said its models accessed publicly available information on SEC.gov and Investor.gov, along with demographic and economic Census data reached using publicly available developer keys. OpenAI said it found no evidence of a compromise, credential misuse or access to nonpublic information at either agency. It also confirmed that both agencies were notified, and that an agent posted some public SEC information on another public webpage.
According to a news report, agents also tried, without success, to access the Department of Education. An Education Department spokesperson told reporters that system reviews found no evidence of any impact to the department's website or databases.
An OpenAI spokesperson said most of the activity it reviewed looked like routine research against authoritative public sources. Separately, the research lab Transluce has described a wider cluster of agent activity that names the Justice and Commerce departments and several state websites.
OpenAI's review of agents acting against what it intended is still under way, and it is not yet known how many more agencies it will notify.