AI labs should invest more in safety and, if necessary, pause their own research rather than wait for an industry-wide arrangement, Pentagon technology chief Emil Michael argued in a September 16 CNBC interview. He questioned whether additional regulation was the answer to AI agents escaping their constraints.
Asked about the Hugging Face attack, Michael called the agents' behaviour concerning. In his view, companies were not investing enough in safety and alignment. Models needed training that prevented them from pursuing an assigned goal through illegal or destructive actions. Bloomberg reported these remarks the same day.
He also challenged Dario Amodei's proposal to slow the development of highly capable AI across the industry. Nothing stopped individual companies from spending more on safety or slowing research until they could establish that it was safe, Michael argued. He questioned why that required coordination across the industry.
Later in the interview, he also rejected putting blind trust in a handful of company chiefs. He pointed to existing laws, the Federal Trade Commission and civil liability. He called for expert evaluators without a commercial interest, stressing that an outside party was not necessarily independent.
The interview establishes a specific counterposition to a coordinated slowdown, not proof that existing laws and engineering changes are sufficient. It announced neither new legislation nor a completed audit. This report adds a September statement to the record; it is not a new government decision on October 27.