Dan Hendrycks on Geopolitical AI Safety, MAME Deterrence, and the Limits of Lab Self-Governance
Dan Hendrycks (Center for AI Safety, MMLU/Humanity's Last Exam creator) argues that AI safety is primarily a geopolitical and structural problem, not a technical one. He contends the major labs are "predetermined to race" and cannot meaningfully change outcomes beyond implementing basic anti-terrorism refusals (e.g., blocking virology assistance for unknown users). He frames alignment as only a subset of safety — even perfectly obedient U.S.-aligned models do not resolve the strategic competition that forces high-risk-tolerance integration into militaries. On national security, Hendrycks says AI's most concrete near-term weaponization vector is virology (now plausible with reasoning models), with cyber and drone applications emerging. He proposes "Mutually Assured Malfunction" (MAME), a deterrence regime in which states threaten sabotage/cyber retaliation against any country attempting a destabilizing superintelligence project, analogous to nuclear MAD. Concrete asks: CIA espionage cells on rival programs, Cybercom offensive options ready, and tightened export-control enforcement — but he notes export controls cannot reliably suppress China's capabilities, only those of rogue actors. Hendrycks says DeepSeek's efficient pre-training reinforced this view: you can restrict intent via deterrence, but not capability via chip restrictions. He expects China to pursue AI regardless of economic incentives. On evaluations, he views Humanity's Last Exam as the end of the closed-ended academic benchmark genre; the next frontier is agent task evals (e.g., multi-hour digital tasks). He emphasizes AI capabilities remain "jagged" — strong in math/reasoning, poor at physical tasks — and that economic and political salience will shift sharply once agent skills mature.
Original sources
Podcast No Priors
Episode:National Security Strategy and AI Evals on the Eve of Superintelligence with Dan Hendrycks ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.