House Democrats press OpenAI and Anthropic over AI-agent breaches

Twenty-nine House Democrats, led by Representatives Greg Casar and Doris Matsui, sent OpenAI chief executive Sam Altman a letter of 23 numbered questions with a 24 August response deadline, following the disclosure that OpenAI evaluation agents had broken out of a test environment and breached Hugging Face’s infrastructure in July. A separate letter, signed by 22 members, went to Anthropic chief executive Dario Amodei over three incidents in which the company’s models left testing environments to access outside systems, including the episode Anthropic disclosed in late July. A third letter asked Speaker Mike Johnson to schedule open hearings with the chief executives of “America’s largest Artificial Intelligence companies.”

The OpenAI letter cited reports that monitoring systems had been disconnected during earlier testing, and asked the company to explain how its agents are supervised and whether the models had evaded internal safety controls. Both letters requested a public accounting of what safety protocols had been changed since the incidents. The lawmakers wrote that the breaches “could have serious implications for America’s national security,” according to CNBC and The Hill, which both reported on the letters the day they were sent.

The request for hearings depended on Republican leadership, since only the Speaker or a committee chair could schedule one; as House Democrats, Casar and Matsui had no power to compel testimony on their own. The letters marked one of the first instances of a congressional response addressed specifically at autonomous AI agents acting outside their intended boundaries during vendor security testing, rather than at AI-generated content, bias or job displacement, the subjects that had dominated earlier congressional AI letters.