What happened
On 28 September 2026 Rep. Ro Khanna (D-CA) announced the 'Human Control Over AI Act,' a summary of which was released to CNBC. The bill would ban recursively self-improving AI and models that autonomously modify their own objectives/containment/shutdown controls until federal guardrails are approved; create a new federal agency to license training/deployment, require embedded independent auditors at frontier labs, set sandbox/air-gap/kill-switch standards, regulate frontier AI chips; impose strict liability and criminal penalties (including 'crimes against humanity' for AI deployment destroying civilian populations, and criminal penalties for employees who disable safeguards); and mandate liability insurance plus enforceable export controls toward China.
Why it matters
It is the most comprehensive US federal AI-safety proposal to date and reframes the week's debate: rather than voluntary White House accords, it proposes binding licensing, strict liability and criminal exposure for frontier developers — a roadmap that would reshape deployment, testing and insurance costs for any lab operating in the US if it gains traction after the midterms.
Action needed
Frontier labs and their counsel should model the impact of licensing/approval gates, embedded-auditor requirements and liability-insurance mandates on release cadence and agent testing; track committee action post-election. Insurance, audit and chip-supply-chain stakeholders should assess the proposed certification roles.