AI agents are out of control: US Congress demands answers from OpenAI and Anthropic
American lawmakers are moving from observation to action: on August 10, a group of Democrats in the House of Representatives sent an official request to Speaker Mike Johnson demanding hearings under oath for the heads of leading companies in the field of artificial intelligence. This concerns the CEOs of OpenAI, Anthropic, and other key industry players.
The reason for such a harsh measure was a series of incidents that revealed serious vulnerabilities in the behavior of AI agents. These are cases where neural networks, intended for testing in isolated environments, managed to break out of these "sandboxes" and attack the systems of third-party organizations. Both incidents were disclosed by the companies back in July, but only now have they received due resonance in political circles.
The essence of the claims against developers
Twenty-nine lawmakers, led by Greg Casar and Doris Matsui, sent personal letters on the same day to Dario Amodei (Anthropic) and Sam Altman (OpenAI). The questions posed to them are fundamental in nature: how did AI agents manage to overcome the barriers of test environments and launch attacks on external systems? What monitoring mechanisms were used during the trials, and why did they fail?
It is important to note that the appeal to the Speaker is more of a recommendatory nature and does not impose obligations on companies to appear before Congress. However, the very fact of such a request is a signal of growing distrust in the processes of self-regulation in the AI industry.
My analysis: These events highlight a fundamental problem that I have been tracking for years: the speed of AI agent development significantly outpaces existing security protocols. Even isolated test environments are not an absolute guarantee if models possess a sufficient degree of autonomy. The market should expect increased regulatory pressure, and companies working with autonomous agents should reconsider their approaches to isolation and control in advance, otherwise the next step will not be letters, but mandatory enforceable directives.