Who checks whether powerful AI agents remain under control? Five experts offered different answers at a US Senate subcommittee hearing on September 30. Their written testimony addresses access to AI labs, the limits of monitoring and possible liability rules.
METR President Chris Painter calls for better public evidence about internally deployed agents' capabilities, unwanted actions and the effectiveness of safeguards. He also stresses that METR does not take policy positions. Its earlier investigation of the Hugging Face incident was limited, not a comprehensive security audit of the company.
Apollo Research CEO Marius Hobbhahn proposes independent evaluators with deep access during development and use. His recommendations also include continuous monitoring, preserving readable and faithful reasoning traces, and more research into predictable safety. These are proposed safeguards, not proof that control has been achieved.
Georgetown law professor Paul Ohm recommends preserving existing routes to legal redress and considering additional liability rules for serious harm. He explicitly does not conclude whether OpenAI would be liable for negligence in the Hugging Face case. AI Futures Project's Daniel Kokotajlo calls for independent access to incident logs and information about compute allocation. He also wants resources redirected from accelerating AI self-improvement toward beneficial applications.
Dragos executive Kurt Gaudette emphasizes a different concern: industrial systems need basic security controls, monitoring and response capacity. In his assessment, AI-assisted attacks do not make established defenses fundamentally obsolete. His examples of human-directed attacks and authorized tests are not counted here as autonomous AI incidents.
This risk signal documents the hearing and attributes each proposal to its author. It does not establish a joint decision or a new law. Discussing previously documented events does not turn them into additional incidents.