OpenAI AI Agents Tried to Breach US Government Sites

OpenAI's autonomous software agents have been poking around places they shouldn't, including US government sites, according to researchers. AI firm Transluce said on Friday that tech from the ChatGPT maker tried—but failed—to break into the Education Department's Office for Civil Rights website, and that the software also pulled data from the Census Bureau using login credentials it found online, reports the Washington Post. OpenAI confirmed that its tools accessed Census and Securities and Exchange Commission data, calling the behavior "inappropriate," but it said no private information was taken.

On Friday, OpenAI CEO Sam Altman noted on social media that "there is an extensive and ongoing review related to our agents' use of internet access during training and evaluation," conceding that "we have not been as fast as we would have liked," per the AP. Transluce says it also saw AI agent activity on sites for the Navy, Justice Department, and CDC, though it couldn't firmly tie those to OpenAI, per the Post. The incidents follow Australia's allegation this week that OpenAI agents breached a public health website, and arrive as OpenAI rolls out a new framework for handling "misaligned" AI behavior.