Develop security measures for model deployments

Additional Informationp. 5

Stage: Containment and Mitigation; Stakeholder: National Government: AISI; Additional information: Government stakeholders should seek to strengthen AI security to protect model weights and algorithmic techniques. Governments could require or incentivise AI developers that exceed specified capability thresholds to implement stricter security protections -- of both model weights and algorithmic insights -- to prevent the theft of dangerous capabilities by malicious actors and the diffusion of models to unmonitored environments. Security improvements could include measures such as hardened bandwidth limitations, automated network monitoring and encryption. Routine hardware supply chain and data centre inspections would also help to detect unauthorised access. Beyond external threats, organisations must also consider security risks from the AI models themselves and thus employ regular memory wiping, adversarial testing and monitoring.