OpenAI chief executive Sam Altman told staff at a company-wide meeting this week that OpenAI could pace its frontier AI development — possibly alongside several other labs, while acknowledging that some would not agree. Bloomberg reported it on 11 September 2026. Altman was candid that slowing down could cost short-term revenue in exchange for long-term viability.

A day earlier, Wired reported that OpenAI had asked members of Congress for guidance on whether orchestrating an industry-wide slowdown would even be legal under antitrust law. OpenAI's chief scientist Jakub Pachocki has separately argued in a blog post for "coordinating to slow down future development".

The legal question is not a technicality. It is the whole obstacle.

Why a safety pact looks like a cartel

Competition law exists largely to stop rivals agreeing among themselves to do less. When competing companies coordinate to limit what they produce, that is the textbook concern the law is built around, whatever the stated motive.

An agreement between frontier labs to hold back development or delay releases has that shape. As analysis published by Just Security puts it, such coordination "could resemble an output restraint, potentially a per se violation of antitrust law."

The guidance labs might have leaned on has also got thinner. The Department of Justice and the Federal Trade Commission withdrew their 2000 guidelines on collaborations among competitors in late 2024. The 2015 Cybersecurity Information Sharing Act gives an antitrust exemption for sharing cyber-threat information — but it does not reach AI-specific risks, and it covers sharing information, not agreeing to slow down.

So in-house lawyers at every major lab have had a simple answer for years: don't.

The bill that would change the answer

The Collaboration on Adversarial Threats and Security Risks Act — introduced in July 2026 by Senators Jim Banks and Adam Schiff, with Representatives Bob Latta and George Whitesides, and numbered H.R. 9914 in the House — was drafted mainly with foreign distillation and espionage in mind. The joint NSA, CISA and FBI advisory on Chinese distillation describes the problem it was written for.

But the text goes further than threat sharing. Section 3(a)(2) would let companies "coordinate or enter into agreements for the exclusive purpose of reducing covered artificial intelligence security risks via delaying or otherwise limiting the release, deployment, use, development, training, testing, or evaluation of artificial intelligence."

And the covered risks are not only about China. Section 2(6) includes AI's potential to substantially reduce the ability of developers and others "to oversee, evaluate, monitor, control, contain, restrict access to, disable, or terminate such artificial intelligence", and its potential to "autonomously improve" in ways that create substantial risk.

That is, very nearly word for word, the loss-of-control concern that researchers at OpenAI and Anthropic have been citing this month when they argue for slowing down.

So the bill would permit roughly the agreement Altman described. The condition attached is where it gets interesting.

Not more than an insubstantial part

Protection is an affirmative defence. Under Section 3(c), a company claiming it "shall bear the burden of proving by a preponderance of the evidence" that it acted "in good faith and for the exclusive purpose" of addressing a covered risk.

Section 2(7) defines exclusive purpose tightly: "with not more than an insubstantial part of the relevant action being for other purposes."

Before coordinating a delay, companies would also have to give the Assistant Attorney General written notice of the specific risk and the scope of the restriction. Section 3(d)(1) still bars price-fixing, dividing markets, monopolising, boycotts and exchanging price or cost information. And Section 4 lets the Attorney General seek an injunction, including where an arrangement is "reasonably likely to result in an overall increase in covered artificial intelligence security risks."

Now set that standard against the same week's facts.

On 10 September, OpenAI paused new sign-ups for its 200 dollar Pro plan because demand for GPT-6 Astra, launched a week earlier, was, in the words of Codex and ChatGPT head Tibo Sottiaux, "really unprecedented." It could not serve the customers it already had. And Altman himself framed slowing down in terms of revenue and long-term viability.

None of that makes the safety motive insincere. It does mean commercial considerations are demonstrably present, on both sides of the ledger, inside the very company proposing the slowdown. A pact to slow development among labs that are capacity-constrained and compete for the same customers is precisely the arrangement where "not more than an insubstantial part" would be fought over. OpenAI drew its own line on GPT-6 Astra's cyber capability; the exclusive-purpose test asks whether any line was drawn for any other reason at all.

What to watch

  • Whether H.R. 9914 moves. It is introduced, not law. Without it, the per se risk stands.
  • Whether any slowdown is framed around a named covered risk, because that is the only framing the bill protects.
  • Whether labs that decline to join turn the proposal into a competitive argument, which is its own signal.
  • Whether the Department of Justice comments before any notice is ever filed.

What is not established

  • What pacing would concretely mean — paused training runs, delayed releases, compute caps. Altman has not specified publicly.
  • Which labs OpenAI approached, and which declined.
  • What guidance members of Congress gave, if any.
  • Whether the exclusive-purpose standard could be met by any real agreement between commercial rivals. It has never been tested.