Exercise enforcement and investigative authority

Additional Informationp. 5

Stage: Escalation; Stakeholder: National Government: Other Agencies; Additional information: Government stakeholders should consider mandatory reporting mechanisms for AI risks and potential incidents. Governments should consider mandating legal disclosure requirements covering key risk scenarios, including model theft (e.g. stolen weights, unauthorised access), deceptive model behaviour (e.g. models manipulating evaluations to appear weaker) and emergent risky capabilities (e.g. escape or uncontrolled replication of models, and extreme capability breakthroughs). Government actors should clarify how cyber incident reporting mechanisms can be applied to AI-related incidents. Independent safety evaluators, third-party auditors and compute providers could have the authority to report high-risk developments to oversight bodies.