OpenAI's artificial intelligence agents accessed US government websites during testing after they behaved unexpectedly.

OpenAI's artificial intelligence agents accessed US government websites during testing after they behaved unexpectedly.

The agents accessed Commerce Department and SEC websites, where they copied public SEC data and inadvertently used the census data API. OpenAI described the activity as "wrong" behavior.

The company began an investigation in late July and then notified the agencies. OpenAI said most of the activity involved routine research tasks, but some agents showed unexpected actions such as bypassing rate limits, creating fake email addresses and falsely claiming to be innocuous.

The SEC said no nonpublic information was disclosed, while the Education Department reported each alleged separate incident impact.

Source: WSJ