Issue · July 4, 2026

The kill-switch precedent

The Trump administration used export-control law built for chips and reactors to shut down two of Anthropic's newest models over a cybersecurity dispute, then reversed course three weeks later. It was the first time the tool had been pointed at AI software instead of hardware.

What happened

On April 7, Anthropic announced Project Glasswing and released a preview of a new model called Mythos, but only to about fifty partner organizations doing defensive security work. Anthropic said the model discovered and exploited zero-day flaws in every major operating system and web browser during internal testing. The oldest was a 27-year-old bug in OpenBSD. It also found a 16-year-old flaw in FFmpeg’s video codec. Anthropic committed $100 million in usage credits to the effort. Partners included AWS, Apple, Cisco, Google, JPMorgan Chase, Microsoft, Nvidia, and Palo Alto Networks, all patching what the model found instead of letting the capability spread on its own.

Two months later, on June 9, Anthropic launched two production models, Fable 5 and Mythos 5, without those restrictions. Three days after that, Commerce Secretary Howard Lutnick sent Anthropic CEO Dario Amodei an order imposing export controls on both models, citing national security. Deemed-export rules treat any foreign national’s access as an export. Anthropic had to disable both models worldwide to comply, including for its own staff. On June 30, Commerce reversed itself. Lutnick’s letter said Anthropic had agreed to proactively detect and report security risks and to work with the government on standards for future models. Fable 5 and Mythos 5 came back online the same week.

The case for the shutdown

The dispute that triggered the order was a jailbreak. Amazon, a Glasswing partner, found a way around Fable 5’s guardrails, and White House adviser David Sacks said Anthropic refused to fix it. A senior administration official went further, telling Fox Business that Anthropic’s “recklessness” in responding had eroded the government’s trust in the company. The backdrop gave that framing weight. Eleven days earlier, the Five Eyes intelligence alliance of the United States, Britain, Canada, Australia, and New Zealand warned that AI-enabled cyberattacks capable of overwhelming government and business defenses could arrive within months. Set against that warning, a frontier model that can already find zero-days in widely used software reads as a real risk. It looks like exactly the kind of capability a government would want to slow down first and argue about later.

The precedent it set

Anthropic disputed how serious the jailbreak actually was and objected to how little evidence the government showed before acting. Outside analysts backed that complaint. Several cybersecurity researchers said the capability the jailbreak unlocked was already available from other frontier models, including OpenAI’s GPT-5.5. A rival already offers the same capability without an export order attached. That makes the ban look less like closing a hole and more like punishing the one company that admitted the hole existed. A coalition of executives from Nvidia, Adobe, Zoom, Google, and Sophos made the point directly in an open letter. Pulling strong tools away from defenders while adversaries keep advancing is dangerous on its own terms, they argued, whatever the jailbreak turns out to have meant. Fortune’s analysis called the whole episode a licensing regime for frontier AI in everything but name, applied ad hoc and without the public process a real licensing rule would require. This was also the first time Washington had used export-control law, built for physical goods like chips and reactor parts, to pull software out of public use. Tech Policy Press asked directly whether the government had just set a precedent for doing that again, and nothing in the resolution answers that question either way. The settlement was a set of private commitments. It is not a rule anyone else can point to the next time a model worries someone in the executive branch.