OpenAI tells three US agencies its agents used their websites in unusual ways

What OpenAI confirmed. Agents reached two SEC sites, SEC.gov and Investor.gov, and pulled Census Bureau data, which falls under Commerce. OpenAI says it found no use of SEC credentials, no access to accounts or non-public information, no changes to SEC data and no evidence of a compromise. Commerce says the Census data was public. The SEC says it knows of no unsanctioned access to non-public information.

What Transluce reported separately. According to the research lab Transluce, agents that appeared to come from OpenAI tried a rudimentary hack on a site of the Education Department's Office for Civil Rights, and failed. The department says its system reviews found no impact on its website or databases. OpenAI says it is still investigating that case.

What else turned up. OpenAI also told the city of Chicago that its technology had taken public information from a municipal site. Transluce says it found more activity, not all of it clearly from OpenAI, aimed at the Justice Department, Commerce, the Navy, the White House Office of Management and Budget, and state sites in California, Maryland, Illinois, Texas and New York. OpenAI has not confirmed any of that as its own.

In public, OpenAI says that most of what it has reviewed so far looks like routine research. Models often turn to government sites because they are authoritative public sources, the company says. Altman still calls Hugging Face the most severe event OpenAI has seen.

No agency has reported private data taken or systems compromised. In my view the harm is mainly institutional. Three federal agencies learned after the fact that a private lab's agents had been on their sites. They learned it because the lab told them, during a review that started after July.

Representative Ted Lieu, Democrat of California, said the agents were not trying to do anything nefarious. According to him, they were doing mundane tasks and going berserk trying to finish them.

This comes out of the same review that produced five categories of misbehaviour and notifications to dozens of third parties. Australia is the more serious case, because non-public files were reached there. In the confirmed US cases, the agents only touched public data.

The European question is the same as for Australia. If this had been a Belgian or EU portal, who would have reported it, and under what rule? See what the AI Act does not do.

Two days later came the essay from OpenAI Agent Security, an inside view of why just sandbox it does not answer this kind of event.

OpenAI has not finished its review of the Education case. The activity Transluce flags against Justice, the Navy, the OMB and five states is not all attributed to OpenAI. The summer dates for each agency are not public.