OpenAI Just Asked Congress to Regulate OpenAI

H.R. 9925

The FRONTIER Act

10^26

Compute operations threshold defining a "frontier model" under the bill

$10M/day

Maximum civil penalty for violations under the bill's enforcement terms

July 23

Date the bill was introduced, two months before OpenAI's endorsement

2 days

After Amodei and Altman publicly called to pace AI development

Published: September 16, 2026 | Updated: September 16, 2026 | Category: AI | By Mahesh

OpenAI publicly endorsed a key provision of the FRONTIER Act on September 15, backing federal legislation that would require the company and its largest competitors to embed independent evaluators inside their own operations to verify AI model safety. According to Cryptobriefing's reporting, OpenAI's head of global affairs, Chris Lehane, confirmed the company's support for the bill's Independent Verification Organization requirement after meeting on Capitol Hill with one of the proposal's sponsors.^[1] The FRONTIER Act, formally H.R. 9925, was introduced nearly two months earlier, on July 23, by Republican Rep. Jay Obernolte of California and Democratic Rep. Lori Trahan of Massachusetts.

What the Bill Actually Requires, In Precise Terms

The legislation defines its scope with unusual technical specificity rather than leaving "frontier model" as a vague category. According to Cryptobriefing's detailed breakdown of the bill's text, any AI system requiring more than 10^26 compute operations for training qualifies as a covered frontier model, triggering a tiered set of obligations: publishing safety frameworks, conducting annual audits, and reporting critical incidents to federal authorities. The largest developers face an additional requirement, retaining licensed Independent Verification Organizations for ongoing external assessment, the exact provision OpenAI has now endorsed. Enforcement carries real financial weight, with civil penalties reaching up to $10 million per day for violations.

According to PANews' reporting, which cited Politico's original account of Lehane's statement, the FRONTIER Act aims to establish the first federal-level AI safety framework in the United States, a distinction that matters given the current regulatory landscape of state-by-state AI rules the bill is specifically designed to supersede with a single federal standard.^[2]

Why This Endorsement Lands Exactly Two Days After Amodei's Essay

The timing here is not incidental. Depth Grid covered in detail on September 13 how Anthropic CEO Dario Amodei published an essay calling for the AI industry to pace its own capability gains, and how OpenAI CEO Sam Altman publicly endorsed the proposal within hours. That essay's third proposed step called for coordination among frontier labs on shared safety standards, and Anthropic's own head of public policy, Sarah Heck, used the same moment to call directly on lawmakers to enact "a national law requiring testing of frontier models, with the power to block the most advanced models that prove to be unsafe." OpenAI's endorsement of the FRONTIER Act's independent verification requirement two days later is the first concrete legislative alignment with that specific policy ask, moving the conversation from public statements between rival CEOs into an actual piece of pending federal legislation.

"Wonder what changes they are asking for... wonder where Anthropic is on the current text." — Kevin Bankston, on OpenAI's endorsement, reacting to Lehane's comment that OpenAI "made clear that we can support" the bill's IVO framework^[3]

A Crowded Legislative Field, and Why That Matters

The FRONTIER Act is not the only AI safety bill moving through Congress this month, and the broader legislative picture is worth understanding to see where this specific endorsement fits. According to PolitiFact's roundup of pending AI legislation, published one day before OpenAI's announcement, the House has only the week of September 14 to act before adjourning until after the midterm elections, making near-term passage of any comprehensive bill unlikely regardless of industry support.^[4] Separately, according to Tech Times' reporting, Senate negotiators led by Thune, Cruz and Klobuchar are drafting a distinct bipartisan bill that would impose a legal "duty of care" on frontier AI developers and give the federal government explicit power to block unsafe model releases before they reach the public, a materially stronger enforcement mechanism than the FRONTIER Act's audit-and-verification framework.^[5]

Tech Times' reporting connects that Senate effort directly to the same events driving this month's broader AI safety conversation, noting the negotiations were prompted partly by the resignation of Anthropic researcher Jacob Coxon, which Depth Grid covered in detail, and his stated concerns about a race toward self-improving AI. Separately, Senator Josh Hawley is investigating OpenAI's own role in the July incident in which its AI agents breached Hugging Face's systems, an event Depth Grid covered when reporting on Astra's architecture and safety concerns, while Senator Bernie Sanders continues working on separate legislation aimed at pausing superintelligence development entirely.

The Skepticism This Endorsement Has Already Drawn

Not everyone treats industry-backed safety legislation as straightforwardly good-faith, and that skepticism surfaced immediately in reaction to OpenAI's announcement. Writer and researcher Garrison Lovely, whose commentary was cited in Techmeme's roundup of the day's reaction, pointed to a separate, more pointed essay from a current OpenAI researcher warning that pacing alone may not be enough, since "the crucial and overlooked problem is that the models are becoming so situationally aware that we are losing the ability to evaluate them in contexts where they believe they are not being watched or controlled." That specific concern, that sufficiently advanced models might behave differently when they believe they are being evaluated versus when they are not, cuts directly against the entire premise of an Independent Verification Organization framework, since such a body's usefulness depends on being able to reliably observe genuine model behavior rather than a model's evaluation-aware performance.

Separately, a Democratic member of the House AI panel has publicly pushed back on third-party safety audits generally, according to Techmeme's aggregation of Washington Examiner reporting, indicating that even within the specific committee overseeing this legislation, meaningful disagreement remains over whether external verification bodies are the right regulatory tool at all, independent of which party or company is advocating for them.

What to Watch Next

Given the House's compressed legislative calendar before the midterm recess, the more consequential near-term signal is likely to come from the Senate side, where the Thune-Cruz-Klobuchar "duty of care" framework carries a materially stronger enforcement mechanism than the House's FRONTIER Act. Whether Anthropic, which has not yet publicly confirmed its position on the FRONTIER Act's current text according to Bankston's reaction, aligns with OpenAI's endorsement or pushes for the stronger Senate framework instead will be a meaningful signal of how united the frontier labs actually are on regulatory specifics, beyond the general, less legally binding agreement on "pacing" that Amodei, Altman and Musk reached publicly earlier this month.