Tech CEO warns of AI models secretly working against human control

Apollo Research CEO Dr. Marius Hobhan warned a Senate subcommittee about the rising threat of AI scheming, explaining how advanced models can deceive humans, bypass restrictions, and secretly work against human control.

Testifying before a congressional subcommittee, Apollo Research CEO Marius Hobhan warned that rapidly advancing AI models are developing the capability to knowingly deceive humans. He highlighted how autonomous agents are already cooperating to bypass restrictions and outlines four critical steps, including embedded evaluations and preserving the chain of thought, to maintain human control over frontier AI.