U.S. curbs Anthropic’s Mythos after suspected China-linked access

On June 13 the Commerce Department imposed export controls that led Anthropic to suspend global access to its cybersecurity models, Mythos and Fable 5. U.S. officials reported a China-linked group had obtained access to Mythos and raised concerns about the potential to replicate the model through a process known as distillation.

Anthropic removed both models from public and commercial use because its systems cannot verify a user’s nationality on a per-API-session basis in real time. Prior to the suspension, access had been limited under Project Glasswing to about a dozen named corporate partners and more than 40 additional organizations for defensive purposes; those arrangements were paused for any parties not explicitly approved by the U.S. government.

Officials have not disclosed which organization accessed the system, how the access happened or how the administration learned of it. The suspension remained in effect as of June 14, and Anthropic has not provided a timeline for restoring service.

Anthropic launched Claude Mythos Preview in April 2026 as a model designed to automate offensive and defensive cybersecurity tasks. Company documentation and testing showed the model could identify thousands of high-severity zero-day vulnerabilities across major operating systems and browsers, reproduce vulnerabilities with 83.1% accuracy on the CyberGym benchmark, find a 27-year-old remote crash vulnerability in OpenBSD and chain Linux kernel exploits to achieve full privilege escalation. Anthropic warned that the model could accelerate the pace of vulnerability discovery.

Fable 5, released in early June, is a consumer-facing variant built with guardrails intended to limit advanced cyber use. White House AI adviser David Sacks posted on social media that Anthropic CEO Dario Amodei had been told Fable 5 was jailbroken, calling the situation ‘serious but easily resolved’ and adding that ‘the ball is in Anthropic’s court.’ Anthropic disputed the severity of the jailbreak, calling it ‘narrow in scope,’ saying similar issues exist in other publicly available models and noting that White House discussions focused on the bypass technique rather than Chinese access concerns.

The loss of access to Mythos affects operators that used the model for defensive scanning. Anthropic’s defensive work using Mythos had identified millions of dollars in potential blockchain exposure and could probe smart contracts, decentralized finance protocols, application programming interfaces and custody systems for weaknesses. Some major exchanges had been in talks to receive defensive access through Project Glasswing; the global suspension removes that pathway for parties not on an approved list.

U.S. officials and security firms have documented prior campaigns that employed AI-assisted techniques against financial institutions and other targets. One campaign linked to Chinese state actors targeted roughly 30 organizations. Some European cybersecurity authorities have warned that Chinese AI providers appear to be limiting public updates while developing offensive and defensive capabilities internally, and China has substantial data-center and compute capacity to support advanced model development.

Regulators have not released detailed public findings about the suspected access. Key questions remain about the amount and type of data that may have been exfiltrated and whether any actors already possess usable copies. The Commerce Department’s export control order restricts broader commercial use of Anthropic’s most advanced cybersecurity models while officials assess next steps.