What Is the FRONTIER Act? H.R. 9925 AI Audits Explained
The short answer
The FRONTIER Act (H.R. 9925) is the first bipartisan federal bill that would legally require the biggest AI companies to submit to outside verification of their safety practices. Sponsored by Rep. Jay Obernolte (R-CA) and Rep. Lori Trahan (D-MA), it licenses Independent Verification Organizations (IVOs) through the Commerce Department, makes the largest labs retain one and get assessed every six months, requires all frontier developers to publish model reports and disclose critical safety incidents, and gives the Commerce Secretary emergency power to suspend a model that poses imminent catastrophic risk.
On September 15, 2026, OpenAI publicly backed the bill’s third-party assessment provision — its first support for a federal audit mandate — while stopping short of endorsing the whole bill. Passage remains uncertain; Roll Call described Congress the same week as “far from erecting guardrails.”
What the bill requires, by tier
| Tier | Threshold (trailing 3 years) | Obligations |
|---|---|---|
| Very large frontier developer | > $5B revenue and ≥ $10B development spend | Publish a frontier AI framework; retain a licensed IVO with full access to operations and records; IVO assesses adequacy against catastrophic risk every six months, reporting to the developer and the Under Secretary simultaneously |
| Large frontier developer | > $50M revenue and ≥ $1B development spend | Publish and follow a frontier AI framework; bring in a competent third-party auditor to certify compliance with that framework (auditor need not be a licensed IVO; framework’s merits not deeply evaluated) |
| All frontier developers | Any | Publish a report on each new frontier model; report critical safety incidents to Commerce |
Two institutional pieces sit on top: a new Under Secretary of Commerce for AI Security to run the IVO licensing regime (setting technical-qualification and independence standards, and stripping licences for biased or shoddy work), and emergency authority for the Secretary of Commerce to suspend development or deployment of a model presenting imminent catastrophic risk. Reporting on the bill cites penalties of up to $10 million per day for non-compliance.
Why the “regulatory market” design matters
The bill deliberately does not write today’s technical safety standards into statute. Instead it borrows the “regulatory markets” idea proposed in April 2023 by legal scholar Gillian Hadfield and Anthropic’s Jack Clark: government sets the objective (acceptable mitigation of catastrophic risk), licensed private verifiers compete on how to assess it, and their licences depend on rigour. The Foundation for American Innovation, which called it “Congress’s best AI bill yet,” argues this keeps the law usable as model sizes, training costs and even the relevant layer of the stack change.
The design also explains why labs can support it. Each company picks its own IVO, the IVO assesses the company’s own published framework, and the government’s role is licensing and emergencies rather than prescribing mitigations. That is closer to financial auditing than to product certification.
From GAAIA to FRONTIER
The bill descends from Obernolte and Trahan’s Great American AI Act (GAAIA), a nearly 300-page discussion draft whose sweeping federal preemption of state AI law drew broad criticism. The FRONTIER Act prunes that: preemption applies only to new state rules on third-party auditing, incident reporting and risk disclosures for frontier models, leaving states their authority over deployment and use. Co-sponsors at introduction include Reps. Houchin, Peters, Scott Franklin and Subramanyam. Consumer group Public Knowledge called it “the best proposal (so far)” for reining in frontier AI.
OpenAI’s position, precisely
OpenAI’s Chris Lehane expressed support at a September 15 roundtable with lawmakers for the provision that an independent organisation must determine a company’s safety protocols are sufficient to mitigate catastrophic risk. A spokesperson told CBS News that OpenAI has not endorsed the bill in full. Lehane’s September 9 post, “AI policy window,” had already called on Congress to enact mandatory national AI safety requirements and on leading developers to build shared industry standards; on CBS Mornings on September 16 he said “that begins with us and the companies.”
Context matters: the same week, Anthropic and other labs published commitments to embedded external evaluators, and PauseAI UK’s Joseph Miller told CBS that developers “probably aren’t going to do it without the government stepping in.” Industry support for audits is now the consensus position; the fight is over who licenses the auditors and what preemption costs the states.
How it interacts with California
California’s AI audit law, signed by Governor Newsom on September 9, 2026, created the first US framework for independent third-party AI verification — and, unlike FRONTIER, it reaches any company deploying AI in hiring, insurance or critical services, not just frontier labs. If FRONTIER passed as written, its narrow preemption would override California’s frontier-model verification and incident-reporting rules but leave the deployment-side obligations intact. That overlap is the likeliest flashpoint in committee.
What happens next
The bill sits in Energy and Commerce and Science, Space, and Technology. There is no Senate companion as of September 17, 2026, and the 119th Congress’s calendar is short. The realistic near-term effect is that the IVO/audit model becomes the template that labs’ voluntary commitments, state laws and any eventual federal statute converge on — regardless of whether H.R. 9925 itself reaches the floor.