US orders Anthropic to cut off access to Fable 5 and Mythos 5 over jailbreak concerns
Government intervention hits Anthropic’s newest models
The U.S. government has ordered Anthropic to immediately suspend access to two of its most advanced AI systems, Claude Fable 5 and Claude Mythos 5, citing national security concerns tied to a potential jailbreak.
Anthropic said it received the directive Friday at 5:21 p.m. ET and has complied, while also arguing the government “got this one wrong.” The company said the order forces it to disable access to both models worldwide, not just for foreign nationals, even though the action was framed as an export control measure.
The move lands at a sensitive moment for Anthropic. Mythos 5 is the company’s most capable model and had been kept under tight restrictions because of its reported ability to uncover security flaws in software. Anthropic said it had tested the model across major operating systems and browsers and found vulnerabilities in each. Rather than release it broadly, the company limited access through Project Glasswing, a controlled program involving about 50 vetted organizations including Amazon, Apple, Google, Microsoft and CrowdStrike.
Fable 5, released just three days ago, was Anthropic’s public-facing version of Mythos 5, designed with guardrails to block high-risk topics such as cybersecurity and biology. The company said it believed those safeguards made it suitable for general release, and benchmark testing from Vals AI had already placed it at the top of the public field.
Anthropic said the government’s concern appears to center on a claimed jailbreak of Fable 5. According to the company, officials have only provided verbal evidence of a “potential narrow, non-universal jailbreak” involving prompts that would get the model to inspect a codebase and identify software flaws. Anthropic said that capability is already present in other publicly accessible models, including OpenAI’s GPT-5.5, and is routinely used for defensive cybersecurity work.
The company also argued that its strongest safeguards rely on independent classifier systems separate from the model itself, meaning a jailbreak would not necessarily defeat the underlying protections against dangerous outputs.
Still, the government’s order is a stark sign of how quickly advanced AI is becoming a national security issue — and how much more willing regulators are to intervene when they believe a model may have crossed a threshold.
Sources: