Threaten to sabotage China's AI
Threatened sabotage of China's AI projects. The U.S. threatens, and if needed carries out, sabotage of rival AI projects it sees as destabilizing.
What it does
The U.S. would openly threaten, and be ready to use, cyberattacks or even strikes to wreck a rival's AI project if it looked like a bid for superintelligence dominance. Rivals would be expected to threaten the same, creating a standoff like nuclear deterrence.
Applies to: Rival nations' frontier AI projects.
- The U.S. threatens, and if needed carries out, sabotage of rival AI projects it sees as destabilizing.
Where things stand
The U.S. relies on chip export controls and intelligence collection. There is no declared policy of sabotaging foreign AI data centers, and no agreed red lines or verification with China.
Why it gains ~4.5 months
Successful sabotage slows China directly: about six weeks of lead gained a year in. The range is very wide.
Biggest unknown: Whether either side can see a rival's AI progress clearly enough to deter without false alarms that escalate to war.
Why it raises p(doom) by ~0.1%
Not a safety law. Uses cyberattacks and other covert means to slow China's most dangerous AI projects.
Debated. Its authors argue the threat deters a reckless race. Critics say it invites escalation and retaliation. We lean slightly toward more risk.
The strongest case that it costs more
Supporters argue mutual deterrence kept the nuclear peace, and a credible threat to sabotage could do the same for AI.
The debate
For
- Dan Hendrycks, Eric Schmidt, Alexandr Wang March 2025: define MAIM as deterrence where a bid for unilateral AI dominance 'is met with preventive sabotage by rivals'.
- Dan Hendrycks and Adam Khoja (AI Frontiers) Sept 22, 2025: 'AI Deterrence Is Our Best Option', a reply to critics.
Against
- David Abecassis, MIRI Apr 2025: MAIM lacks monitorable red lines and sabotage only delays; deterrence should target chip chokepoints instead.
- Jason Ross Arnold (AI Frontiers) Aug 2025: MAIM has an observability problem that could trigger false alarms or miss real breakouts.
- RAND Mar 2025: argues MAIM misreads how Cold War deterrence worked, since MAD did not rest on preventive sabotage.
Sources
- Superintelligence Strategy (nationalsecurity.ai): Original MAIM definition and escalation ladder up to kinetic strikes on data centers.
- Superintelligence Strategy: Expert Version (arXiv 2503.05628): Full paper.
- IAPS: Crucial Considerations in ASI Deterrence (Oscar Delaney): Estimates ~25% chance MAIM dynamics meaningfully slow AI.
- AI Frontiers: AI Deterrence Is Our Best Option: Authors' response to Wildeford, Delaney, Abecassis, Rehman, Arnold.
Rough starting points, not precise forecasts. Lead costs assume China doesn't depend on U.S. models, the case least favorable to safety laws, and count 3 years. On the menu you can change every assumption and put in your own numbers. Last priced 2026-09-26.