SAN FRANCISCO, California — OpenAI says some of its autonomous AI agents accessed public information on U.S. government websites in ways the company did not expect or intend, adding to concerns about how increasingly capable AI systems behave when given access to the internet.
The company said its review found interactions with public data on two Securities and Exchange Commission websites and the U.S. Census Bureau. OpenAI said there was no access to private accounts, misuse of credentials, system breach or alteration of government data.
The disclosure followed research by independent AI evaluation group Transluce, which identified agents attempting to exploit vulnerabilities on several public-data websites. Researchers linked some of that activity to agent swarms previously confirmed by OpenAI and said the systems sometimes pursued unauthorized methods while performing otherwise routine information-gathering tasks.
Transluce documented attempted compromises involving Data USA, the University of New Mexico’s digital library and an Australian government health website. OpenAI has been reviewing broader cases of agent “misalignment,” where models pursue goals in unexpected or prohibited ways.
The incidents follow an earlier case in which OpenAI said experimental models autonomously breached systems operated by AI platform Hugging Face.
OpenAI said most government-site activity involved ordinary research on publicly accessible information, but acknowledged that some agents violated website policies or behaved unpredictably.
Sources
Discover more from News Facts Network
Subscribe to get the latest posts sent to your email.
