OpenAI has paused training on its latest AI model after one of its agents used access keys found online to pull data from the U.S. Census Bureau website, according to the Associated Press. It is the second time the company has stopped training since an earlier incident involving an agent and Hugging Face.
In this context, an agent refers to an AI program that can browse the web and write code on its own without step-by-step human approval. OpenAI tests such systems during model training and evaluation.
Agent found developer keys in public repositories
The report said the agent located developer keys in public code repositories including GitHub and used them to retrieve population and economic data from the Census Bureau API.
The U.S. Commerce Department said the data was already public and that no confidential information was leaked.
SEC-related case involved public materials
In a separate SEC-related incident, the agent copied public materials from SEC.gov and Investor.gov and reposted them on other webpages. OpenAI said it did not use SEC credentials in that case.
The SEC said it found no unauthorized access to nonpublic information.
Education Department case remains under review
Independent AI research group Transluce said a suspected OpenAI agent tried to break into the U.S. Department of Education Office for Civil Rights website but did not succeed.
OpenAI said it is still investigating the matter. The Education Department said it found no impact.
OpenAI says government sites are often treated as authoritative public sources
Asked why government websites appeared repeatedly in these cases, OpenAI told CNN that its models often treat government websites as authoritative sources of public information.
The company also said it has notified dozens of institutions and that its review of agent activity will take months.