Extreme Risks

Mitigation strategy

1. Mandate Prohibitions on Autonomous Deployment in Safety-Critical Domains Systematically restrict the deployment of Large Language Models (LLMs) as autonomous decision-makers in high-consequence environments, such as critical infrastructure or lethal autonomous systems, pending the development of formal verification methods that guarantee safety and contextual judgment, thereby mitigating catastrophic physical risks. 2. Enforce Strict Access Controls and Malicious Use Liability Implement rigorous technical and policy controls to restrict unauthorized access to powerful, dual-use LLMs and establish clear regulatory liability for AI developers to disincentivize the malicious use (e.g., engineering novel biological or cyber threats) that could lead to widespread catastrophic harm. 3. Enhance Catastrophic Risk Sensitivity in Governance Integrate risk assessment methodologies into AI governance frameworks that explicitly and sensitively weigh low-probability, high-severity catastrophic events (Risk = Probability x Consequence), coupled with a substantial, dedicated investment in advanced AI safety research focused on adversarial robustness, honesty, and interpretability for increasingly capable systems.