U.S. Orders Anthropic to Shut Down Two Flagship AI Models

The U.S. government issued an emergency directive Friday ordering Anthropic to immediately disable two of its most advanced AI models worldwide, marking an unprecedented federal intervention in commercial AI deployment.

Anthropic confirmed it received the order at 5:21 pm ET on Friday and has complied, shutting off access to Claude Fable 5 and Claude Mythos 5 for all users globally. The company's other models remain operational.

The models at the center of the shutdown

Mythos 5 represents Anthropic's most capable AI system to date. The company previewed the model in early April but never released it publicly, citing its exceptional ability to identify security vulnerabilities across software systems. According to Anthropic, Mythos found flaws in every major operating system and web browser it tested.

Instead of a broad release, Anthropic launched Project Glasswing, a controlled access program that shared Mythos with approximately 50 vetted organizations including Amazon, Apple, Google, Microsoft, and CrowdStrike for defensive cybersecurity work.

Fable 5, released just three days before the shutdown, was positioned as a commercially viable alternative—essentially Mythos with guardrails blocking responses in high-risk domains like cybersecurity and biology. Independent benchmark tests from Vals AI rated it as the most capable publicly available AI model at launch.

Why it matters

This case reveals a potential regulatory paradox: companies that publicly emphasize AI safety risks may invite more aggressive government oversight than competitors who downplay those same concerns. If the government applies this standard consistently across the industry, it could fundamentally reshape how AI companies communicate about model capabilities and establish a new precedent for federal intervention in commercial AI releases.

The jailbreak claim

The government framed its action as an export control measure restricting foreign national access. However, Anthropic says officials indicated the real concern centers on a claimed jailbreak of Fable 5—a technique that allegedly bypasses the model's safety restrictions.

Anthropic disputes the severity of this vulnerability. The company says it has received only verbal evidence of what officials describe as a "potential narrow, non-universal jailbreak" that involves prompting the model to analyze specific codebases for security flaws. Anthropic argues this capability level already exists in other publicly accessible models, including OpenAI's GPT-5.5, and represents standard practice in defensive cybersecurity work.

The company maintains that its core safety mechanisms operate through independent classifier systems separate from the model itself, meaning protections against dangerous outputs remain intact even if conversational guardrails are bypassed.

The safety messaging backfire

Anthropic built its brand identity around being the safety-conscious AI company. That positioning now appears to have created regulatory exposure. The company spent months restricting Mythos access while emphasizing its potential dangers—messaging that may have amplified government concerns rather than demonstrating responsible stewardship.

OpenAI CEO Sam Altman criticized this approach in April, telling podcaster Ashlee Vance that Anthropic's handling of Mythos amounted to "fear-based marketing." Altman suggested the strategy of highlighting a model's dangers while offering controlled access resembled selling bomb shelters after announcing you've built a bomb.

Anthropic warned in its response that applying this shutdown standard industry-wide "would essentially halt all new model deployments for all frontier model providers." Both Anthropic and OpenAI are expected to pursue public offerings this year, making regulatory stability particularly consequential for their business trajectories.

These details were first reported by TechCrunch.