A new policy paper from AI safety expert Dan Hendrycks, former Google CEO Eric Schmidt, and Scale AI CEO Alexandr Wang argues that the race for AI dominance is inherently destabilizing, and the only way to keep it in check is deterrence through sabotage. Dubbed Mutual Assured AI Malfunction (MAIM), their framework suggests that any nation racing toward unchecked superintelligence could face cyber or even physical attacks on its AI infrastructure from rivals who see such an effort as an existential threat.
Key Points:
- MAIM draws parallels to nuclear deterrence, proposing that rival states may sabotage each other’s AI projects rather than allow unilateral AI dominance.
- Cyberattacks, insider threats, and datacenter disruptions could be used to disable destabilizing AI projects before they reach dangerous capabilities.
- The AI arms race creates incentives for preventive strikes if one nation appears to be pulling ahead in developing superintelligence.
- A broader AI strategy should include deterrence, nonproliferation, and economic competitiveness to balance security with innovation.
The AI Arms Race Just Got Darker
As the race toward artificial superintelligence heats up, so do fears of an AI-driven geopolitical crisis. A new policy paper authored by Dan Hendrycks, Eric Schmidt, and Alexandr Wang proposes a drastic approach to preventing AI hegemony: sabotage. Their concept, Mutual Assured AI Malfunction (MAIM), argues that just as nuclear deterrence during the Cold War relied on the threat of retaliation, AI superpowers may find themselves forced to engage in preemptive disruption to prevent rivals from gaining a decisive AI advantage.
The authors warn that a state pushing too aggressively toward AI supremacy creates two unacceptable risks: either it loses control of its system, endangering global security, or it succeeds in establishing AI dominance, threatening its rivals’ national survival. In response, adversarial nations might resort to sabotage—whether through cyberattacks, insider threats, or, in extreme cases, kinetic strikes on datacenters. The paper claims that “we are already approaching a dynamic similar to nuclear Mutual Assured Destruction (MAD).”
Deterrence Through Sabotage
The framework suggests that because AI projects are fragile—dependent on compute-intensive infrastructure and vulnerable to cyber exploits—states will likely develop offensive cyber capabilities to disable rival AI programs. The authors outline a potential escalation ladder: intelligence gathering on AI projects, covert cyber interference to degrade training runs, overt cyberattacks on power grids or cooling systems, and, in extreme cases, physical destruction of datacenters. The aim is to create a stable deterrence regime where no state attempts an uncontested AI sprint without expecting retaliation.
However, unlike nuclear weapons, AI capabilities are software-based and decentralized, making deterrence harder to enforce. While a nuclear arsenal is physically limited and relatively easy to monitor, AI research is harder to track, meaning covert development efforts may persist despite attempted sabotage.
Nonproliferation and Competitiveness
Beyond deterrence, the paper argues for strict AI nonproliferation policies to keep powerful models out of the hands of rogue actors. Proposals include:
- Compute security: Tracking and controlling high-end AI chips to prevent smuggling.
- Information security: Protecting model weights from being leaked or stolen.
- AI security: Embedding safeguards to prevent malicious use, similar to biosecurity measures for handling pathogens.
Additionally, the authors emphasize that nations must bolster their AI capabilities through domestic chip manufacturing, military AI integration, and legal frameworks governing AI behavior to remain competitive.
The Big Picture
While acknowledging the transformative potential of superintelligent systems, the paper urges pragmatism over fatalism or denial. "States that act with pragmatism instead of fatalism or denial may find themselves beneficiaries of a great surge in wealth," the authors conclude, suggesting that a measured approach to AI governance could facilitate unprecedented benefits rather than catastrophic outcomes.
They present their "Multipolar Strategy" as a middle path between unregulated AI development and attempts to monopolize advanced AI capabilities through government-led initiatives like a hypothetical "AI Manhattan Project."
The Implications of MAIM
The MAIM framework paints a grim picture of an AI future driven by mistrust and geopolitical maneuvering. It suggests that superintelligence development is unlikely to be a cooperative global effort and instead will be shaped by espionage, cyber conflict, and power struggles akin to the Cold War.
Whether MAIM will emerge as a formal policy doctrine remains to be seen, but the authors make one thing clear: the AI arms race is no longer just about who can build the smartest system—it’s also about who can prevent their rivals from doing the same.