Regulation

Government institutes testing frontier models before release.

The US Center for AI Standards and Innovation has run more than 40 assessments, and all five major US labs had signed testing agreements by 5 May 2026. It has no permanent legal standing.

1. 01SecureWorld: US AI labs and government security reviews

Labs’ own published rules for when to add safeguards or stop.

Anthropic’s RSP v3 removed its pause commitment and made most mitigations conditional, a weakening in 2026. Its risk reports came out in February and August 2026.

1. 01GovAI: Anthropic’s RSP v3.0, what has changed
2. 02Anthropic: Responsible Scaling Policy

Laws that make frontier developers publish safety plans and report incidents.

California’s SB 53 has applied since 1 January 2026 to developers training above 10^26 FLOP with more than $500m in revenue. New York’s RAISE Act was signed in December 2025. Both require critical-incident reporting.

1. 01Cloud Security Alliance: US AI regulation and preemption
2. 02Future of Privacy Forum on New York’s RAISE Act

Legal protection for lab employees who report safety problems.

Senator Grassley’s federal bill had no committee or floor action as of 25 September 2026.

1. 01Deseret News: the AI whistleblower bill

Federal moves to override state AI rules such as SB 53.

A Justice Department task force existed by January 2026, but no suit had been filed and no state law struck down as of April 2026.

1. 01Cloud Security Alliance: US AI regulation and preemption

Regulators letting AI perform a licensed act, such as a prescription renewal or a legal letter, within a narrow sandboxed scope.

Utah has let Doctronic renew prescriptions for 192 chronic-condition drugs since about January 2026, with physician review of the first 250 patients; the SRA authorised the first AI-only law firm, for debt claims up to £10,000, in May 2025.

1. 01Utah's experiment with AI-driven prescription renewals (Stanford Law, Mar 2026)
2. 02SRA approves AI-driven law firm (Legal Cheek, May 2025)