Claude Mythos Preview, a frontier AI model Anthropic built for cybersecurity work, was pulled offline three days after launch by a U.S. national security order that Anthropic implemented globally, taking a defender's tool away from the people it was built to help.
The bug that Mozilla's security team shipped in Firefox 150 was not glamorous. It was a memory handling flaw in the browser's image decoder, the kind of quiet mistake that, left in the wild, becomes a doorway for attackers. The fix was straightforward, and so were the 270 others like it. Mozilla engineers, working with a new kind of partner, walked the Firefox 150 codebase clean of 271 vulnerabilities before the browser ever shipped, a haul the project's security lead described as the largest single-release cleanup in the browser's history.
The partner was Claude Mythos Preview, a frontier AI model Anthropic released on June 9 alongside Fable 5 — a closely related variant with safeguards lifted for security research. Both models were positioned as tools for the defensive side of the field, designed to read code, reason about adversarial behavior, and surface the kind of subtle flaws that human reviewers miss under deadline pressure. Three days after launch, both models were offline for every customer, including Anthropic's own foreign-national employees, after a U.S. national-security order restricted foreign nationals' access to the systems and Anthropic concluded it could not reliably separate customers by location.
The order came from the Trump administration, which told Anthropic to suspend foreign nationals' access to Fable 5 and Mythos 5 over concerns that the same capabilities could be turned against U.S. networks by hostile states or criminal groups. Anthropic, according to Scientific American's reporting on the episode, disabled the models for all customers rather than attempt a narrower cut, on the grounds that the order's scope was effectively global and that the company could not guarantee compliance with a more limited version of the restriction.
That is the dilemma at the center of this story, and it is not an abstract one. The same class of AI that helped Mozilla triage 271 bugs in a single release is the class that security researchers say is already being probed by attackers for use in phishing, exploit development, and vulnerability discovery at scale. Cybersecurity experts quoted in the piece argue that restricting the defensive side removes a tool the blue team was actively using, and that attackers, who do not wait for export-control compliance, will keep building with whatever frontier models they can obtain. The dual-use problem is not a slogan in this story; it is the mechanism.
The defender loss is concrete. Mozilla's 271 fixes in Firefox 150 were the most visible example, but Scientific American's reporting also describes Cloudflare's testing of Mythos for its own security workflows, and broader bug-hunting work that had been built around Anthropic's Opus 4.6 model. None of those defender uses failed on their merits. The tools were working, the workflows were real, and the people running them were the ones the policy was nominally trying to protect.
The stated national-security rationale is legitimate on its face. Adversaries do want frontier AI, and a domestically built model with cybersecurity strengths is a reasonable thing to want to control. The critique is not that the goal was wrong. The critique is that the mechanism was blunt, and that the cost of bluntness fell on the defenders the model was built to help, with no compensating process for distinguishing adversary use from defender use before the model was withdrawn.
That is the governance question the next frontier model is going to run into. Claude Mythos Preview was online for three days. The next system with comparable capability will be online for longer, used by more defenders, and useful to more attackers, and the policy that pulls it will face the same trade-off. The interesting question is not whether frontier AI should be controlled; it is what a control regime would have to look like to keep a useful defensive tool in defenders' hands while denying it to the people it is meant to be denied to. The current answer, in the Mythos Preview episode, was to take it away from everyone. Firefox shipped 271 bugs cleaner for it, and then the model was gone.