Semafor: China-Linked Group Suspected of Accessing Anthropic's Top-Secret AI

The Trump administration ordered Anthropic to restrict its most advanced cybersecurity artificial intelligence (AI) model on June 13, 2026, prompting the company to suspend access worldwide. Semafor reports that a China-linked group was suspected of gaining unauthorized entry to the system.

• Key Takeaways:
• • The Trump administration issued export controls on June 13, forcing Anthropic to pull Mythos and Fable 5 globally.
• • Semafor reports China-linked actors are suspected of accessing Mythos, raising reverse-engineering fears for U.S. security.
• • Crypto platforms, including Coinbase, lose defensive AI access as Anthropic’s Project Glasswing faces an uncertain timeline.

What Happened

Anthropic shut down global access to two AI models, Mythos and Fable 5, following a Commerce Department directive citing national security concerns. Because verifying user nationality per API session in real time is not technically feasible, Anthropic removed both models from the market entirely rather than attempt selective enforcement.

Semafor, in an exclusive report by Reed Albergotti published this week, said the export control decision was partly motivated by suspicions that a China-linked group had accessed Mythos. If the Chinese government obtained the model, U.S. officials feared Beijing could reverse-engineer it through a process known as “distillation,” effectively replicating its capabilities.

It remains unclear which specific organization was involved, how access occurred, or how the White House learned of it.

What Is Mythos

Anthropic launched Claude Mythos Preview in April 2026 as a frontier AI model built around autonomous cybersecurity offense and defense. The company deemed it too dangerous for public release, citing its ability to identify and exploit previously unknown software vulnerabilities at scale.

Key capabilities documented by Anthropic include:

• Autonomous identification of thousands of high-severity zero-day vulnerabilities across major operating systems and browsers
• Discovery of a 27-year-old remote crash vulnerability in OpenBSD, used in critical infrastructure
• Chained Linux kernel exploits enabling full privilege escalation
• 83.1% accuracy on the CyberGym vulnerability reproduction benchmark

Anthropic estimated AI-accelerated attacks could compress what currently takes months of expert work down to minutes. Cybercrime already costs an estimated $500 billion annually worldwide.

Access was restricted to roughly 12 named corporate partners and 40-plus additional organizations through Project Glasswing, a defensive initiative designed to find and patch vulnerabilities before malicious actors could exploit them.

The Fable 5 Jailbreak

A second model, Fable 5, a consumer-facing version of Mythos built with guardrails against advanced cyber use, was released in early June. White House AI adviser David Sacks posted on X that Anthropic CEO Dario Amodei was notified that Fable 5 had been jailbroken but reportedly downplayed the risk and declined to address it.

Sacks said the administration “values Anthropic’s technical capabilities” and described the situation as “serious but easily resolved,” adding that “the ball is in Anthropic’s court.”

Anthropic disputed the severity of the jailbreak, stating it was narrow in scope and involved issues already present in other publicly available models. The company also stated the White House did not raise Chinese access concerns during direct discussions, which focused on Fable 5 and the bypass technique.

Why Crypto Platforms Are Watching

Mythos has direct relevance to digital asset security. Anthropic’s work using the model identified millions in potential blockchain exploit exposure, and the model can autonomously probe smart contracts, decentralized finance ( DeFi) protocols, APIs, and custody systems for weaknesses.

Major exchanges, including Coinbase, had reportedly entered discussions about defensive access through Project Glasswing. The global access suspension halts those arrangements for non-approved parties and removes a significant defensive tool from operators outside the approved partner list. More recently, just before Claude Fable 5 and Mythos 5 were banned, Zcash founder claimed Mythos audited the whole Zcash chain and found no bugs.

Security professionals who relied on the model’s vulnerability scanning capabilities now have no access path. That gap benefits any actor that may have already obtained a copy.

US-China AI Competition

The incident fits a broader pattern. Semafor’s reporting and prior disclosures connect Chinese state-linked actors to earlier attempts at using Anthropic models for cyber espionage. A campaign documented by Anthropic, labeled GTG-1002, linked Chinese state actors to AI-assisted attacks against roughly 30 entities including financial institutions.

Germany’s cybersecurity officials have separately warned that Chinese AI providers appear to be reducing public updates while building Mythos-equivalent capabilities internally. China has the data center infrastructure and compute capacity to pursue that development path.

The suspension of Mythos and Fable 5 access remains in effect as of June 14, 2026. Anthropic has not announced a timeline for restoring service.