Requires independent approval before frontier AI models or major upgrades reach the public.

• Defines covered models through adjustable capability, computing, and public-risk thresholds.

• Requires testing (dependent on AI model) for cybersecurity, biological or chemical access, autonomous actions, deceptive practices, and self-preservation issues.

• Allows regulators to approve, condition, delay, or prohibit deployment based on demonstrated risk.

Prevents AI developers from controlling or compromising the safety reviews they undergo.

• Establishes accreditation, competency, independence, and conflict-of-interest requirements for evaluators.

• Funds reviews through an independent fee pool rather than direct developer payments.

• Requires transparent evaluation methods, regulatory audits, and penalties for falsified findings.

Requires advanced AI systems to be trained around safety, rights, and human dignity.

• Establishes training standards grounded in civil rights, privacy, child safety, truthfulness, and human autonomy.

• Requires independent ethicists, workers, educators, disability advocates, and public-interest experts to review training practices.

• Tests models for manipulation, dehumanization, discrimination, dangerous guidance, and circumvention of established safeguards.

Protects workers who expose concealed AI dangers, misconduct, fraud, or regulatory violations.

• Prohibits firing, demotion, intimidation, blacklisting, and other retaliation against protected whistleblowers.

• Establishes secure and confidential reporting channels independent of company management that are required for every company employee or subcontractor.

• Provides reinstatement, lost wages, damages, legal expenses, and civil penalties for retaliation.

Requires frontier developers to protect AI models from theft, sabotage, or misuse, and to provide clear “surfaces” for model inspection.

• Establish cybersecurity standards for model weights, computing systems, sensitive research, and development infrastructure.

• Requires strong inspection access controls, employee safeguards, supply-chain security, and continuous threat monitoring.

• Mandates independent security audits and immediate reporting of AI breaches or attempted model theft.

Establishes safeguards for AI systems capable of independently taking consequential real-world actions.

• Requires clear human authorization before agents move money, access accounts, or control systems.

• Mandates secure identification, activity records, spending limits, access restrictions, and emergency shutdown controls.

• Requires testing against consumer or operator manipulation, engaging unauthorized actions, and attempts to evade supervision or shutdown.

Creates heightened safeguards for future AI controlling services utilized on public safety and survival systems.

• Establishes regulations on AI in energy, water, communications, transportation, health care, finance, emergency services, and defense.

• Requires continuous human supervision, manual operating options, and reliable fail-safe systems.

• Mandates independent testing, emergency exercises, incident response plans, and recovery procedures.