Warner, Schatz and Kim propose mandatory AI testing and incident reporting

Senators Mark Warner, Brian Schatz and Andy Kim announced proposed AI security legislation on September 24. It would create an AI Safety Board within the Commerce Department and require safety plans, testing access and incident reporting. These are proposed duties, not requirements established by this announcement. Senator Warner · announcement 

Section 3(f) would require developers to provide model access at least 45 calendar days before introduction into interstate or foreign commerce. Sections 3(c)–(d) provide for standards compliance, enforcement by the Commerce Secretary and court injunctions sought by the Attorney General. The text does not set out a general requirement for the Board to approve every model’s release. Section 5(a)(3) would require reporting confirmed incidents within 30 days of confirmation, or within 72 hours for imminent threats. Its scope includes non-public models and activity during development and testing. The proposal also addresses standards governing agents’ authority and access. The proposal supplies a concrete version of the argument for enforceable outside scrutiny: advance access and standards backed by enforcement. That differs from relying entirely on developers’ voluntary evaluations, while leaving the details and effectiveness of the proposed system to be established. Senator Warner · proposed bill