National Security: Superintelligence Strategy

Summary of https://arxiv.org/pdf/2503.05628

This expert strategy document from Dan Hendrycks, Eric Schmidt and Alexander Wang addresses the national security implications of rapidly advancing AI, particularly the anticipated emergence of superintelligence.

The authors propose a three-pronged framework drawing parallels with Cold War strategies: deterrence through the concept of Mutual Assured AI Malfunction (MAIM), nonproliferation to restrict access for rogue actors, and competitiveness to bolster national strength.

The text examines threats from rival states, terrorists, and uncontrolled AI, arguing for proactive measures like cyber espionage and sabotage for deterrence, export controls and information security for nonproliferation, and domestic AI chip manufacturing and legal frameworks for competitiveness. Ultimately, the document advocates for a risk-conscious, multipolar strategy to navigate the transformative and potentially perilous landscape of advanced artificial intelligence.

• Rapid advances in AI, especially the anticipation of superintelligence, present significant national security challenges akin to those posed by nuclear weapons. The dual-use nature of AI means it can be leveraged for both economic and military dominance by states, while also enabling rogue actors to develop bioweapons and launch cyberattacks. The potential for loss of control over advanced AI systems further amplifies these risks.

• The concept of Mutual Assured AI Malfunction (MAIM) is introduced as a likely default deterrence regime. This is similar to nuclear Mutual Assured Destruction (MAD), where any aggressive pursuit of unilateral AI dominance by a state would likely be met with preventive sabotage by its rivals, ranging from cyberattacks to potential kinetic strikes on AI infrastructure.

• A critical component of a superintelligence strategy is nonproliferation. Drawing from precedents in restricting weapons of mass destruction, this involves three key levers: compute security to track and control the distribution of high-end AI chips, information security to protect sensitive AI research and model weights from falling into the wrong hands, and AI security to implement safeguards that prevent the malicious use and loss of control of AI systems.

• Beyond mitigating risks, states must also focus on competitiveness in the age of AI to ensure their national strength. This includes strategically integrating AI into military command and control and securing drone supply chains, guaranteeing access to AI chips through domestic manufacturing and strategic export controls, establishing legal frameworks to govern AI agents, and maintaining political stability in the face of rapid automation and the spread of misinformation.

• Existing strategies for dealing with advanced AI, such as a completely hands-off approach, voluntary moratoria, or a unilateral pursuit of a strategic monopoly, are flawed and insufficient to address the multifaceted risks and opportunities presented by AI. The authors propose a multipolar strategy based on the interconnected pillars of deterrence (MAIM), nonproliferation, and competitiveness, drawing lessons from the Cold War framework adapted to the unique challenges of superintelligence.