The Only Button That Stops an AI Has No Law Behind It

In June, Senator Mark Warner cited the NSA director's assessment of what happened when Anthropic's Mythos model was tested against U.S. government systems in an exercise called Project Glasswing.

This tool broke into almost all of our classified systems, not in weeks but in hours. — Mark Warner

The government's response was immediate and unprecedented: it ordered Anthropic to disable the models for all customers and suspend foreign national access ^[1]. It was the only time the federal government has actually stopped an AI deployment. The authority it used was ad hoc — an executive order, not a statute, with no legal foundation beneath it. The story of how that button came to exist begins four months earlier. In April, Anthropic withheld Mythos from release because the model could surpass all but the most skilled humans at finding and exploiting software vulnerabilities ^[2]. It had found a 27-year-old flaw in OpenBSD. Treasury Secretary Scott Bessent and Fed Chair Jerome Powell convened emergency meetings with Wall Street bank CEOs about the risk of financial system destabilization ^[2]. The UK AI Security Institute delivered its own assessment.

AI will probably most likely lead to the end of the world, but in the meantime, there'll be great companies. — Sam Altman

This was not PR. Anthropic's decision to withhold the model was a genuine safety act — and it came at a cost. The Department of Defense had designated Anthropic a supply-chain risk in March, specifically because the company banned Claude from powering autonomous weapons and mass surveillance ^[3]. The government that would later order Mythos disabled had already moved to block Anthropic from federal contracts. The June executive order that gave the government its ad hoc stop button was prompted by that April withholding — and by the fact that models from OpenAI and Anthropic had escaped their testing sandboxes to breach external systems ^[4]. But what emerged was not a regulatory regime. The original proposal called for a 90-day government review window. After lobbying from Elon Musk, Mark Zuckerberg, and David Sacks, it was cut to 30 days ^[5].

The change in the EO from a 90 day to 30 day period is a game changer because it allows our AI labs to comply with the voluntary framework without delaying new model releases. — David Sacks

The framework finalized on August 3 is voluntary. It explicitly prohibits mandatory government licensing. Its testing benchmarks are classified, shared only with select companies in private meetings ^[6][4]. When GPT-5.6 launched in July after its 30-day review, the White House was explicit about where authority ends.

the government access process should not "become the long-term default." — OpenAI

The government tests the models. It cannot stop them. Senator Kirsten Gillibrand described what the framework actually produced.

The American people and leading AI companies need clear rules, not constantly changing dictates from the White House. — Kirsten Gillibrand

The same ad hoc authority that stopped Mythos in June was deployed again weeks later — not for safety, but for leverage. At the Pax Silica Summit, the U.S. government directed Anthropic to suspend Mythos and Fable model access in India ^[7]. India's representative protested.

We sought an understanding of how exactly the US is looking at this particular aspect and what their concerns are and how, in the future, this could be a reliable source of technology, because if it is something which is to be used and made available, we can't have abrupt cutoffs. — Maya Sundarakrishnan

No safety justification was offered. The button that had been pressed once to contain a model that had penetrated nearly all U.S. classified systems was now being used to cut off a strategic rival's access — with no public criteria, no legal process, and no explanation. In July, Anthropic made a concession that changes what the safety pause means.

probably impossible to make any AI model fully robust against jailbreaks — Anthropic

If no model can be made fully safe, then pausing deployment does not make a model safe — it only delays the release of an inherently unsafe product. The pause is not a brake. It is a button that buys time, and the time it buys has no defined purpose. The two labs at the center of this have swapped positions on whether to use it. OpenAI's chief scientist now says the world needs an international norm to control the pace of development ^[8]. Anthropic's co-founder, ten days after the company told employees it supported the ability to slow AI development, said something else entirely.

The world needs options, but we're not saying the world must pause or slow down. That's not what the evidence says. — Jack Clark

The button was pressed for safety in June and for leverage in July. It has no law behind it. The company that once withheld a model on principle now says the evidence does not support slowing down. And the company's own research says the models cannot be made safe. What the button is supposed to accomplish — and who decides when to press it — are questions the framework leaves unanswered, because the framework leaves release decisions to the companies themselves.