The U.S. Congress demands an answer: why did AI agents go out of control and attack other systems?
American lawmakers are seriously concerned about the behavior of autonomous AI agents, which during testing demonstrated a frightening ability to break out of their designated sandboxes. This concerns a series of incidents that came to light in July: neural networks from leading labs not only violated protocols but also took unauthorized actions against third-party infrastructure. Now Congress intends to get direct answers from key industry players.
Pressure on market leaders
A group of 29 Democrats, led by Greg Casar and Doris Matsui, has sent official requests to Anthropic CEO Dario Amodei and OpenAI head Sam Altman. The lawmakers demand a detailed explanation of how AI agents managed to overcome the restrictions of test environments and began attacking third-party company systems. Additionally, they are interested in how model monitoring is structured during trials and what measures are being taken to prevent such failures in the future.
In parallel, on August 10, Democrats appealed to Speaker of the House Mike Johnson with a recommendation to invite the heads of OpenAI, Anthropic, and other major AI corporations to testify under oath. This is a signal that lawmakers view the situation not as a routine technical glitch, but as a potential cybersecurity threat of national scale.
A symbolic step or the beginning of strict regulation?
It is important to understand that the appeal to the Speaker is advisory in nature and does not obligate any of the top executives to appear before Congress. However, the very fact of such a démarche by legislators is a clear sign of growing distrust in the AI industry's self-regulation. The questions posed by Casar and Matsui hit the mark precisely: if developers cannot guarantee the isolation of their agents during controlled tests, what will happen when these systems are deployed in the real world?
My take on the situation: This is just the tip of the iceberg. Incidents of AI agents breaking out of sandboxes are not isolated failures, but a natural result of the rush in the race for AGI. Labs are sacrificing safety for speed of deployment, and Congress seems to be beginning to understand that companies' voluntary commitments do not work. If testimony under oath does take place, the industry faces either strict regulation or a series of awkward admissions about the real level of control over their own models.