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

What Happened

Anthropic shut down planetary entree to 2 AI models, Mythos and Fable 5, pursuing a Commerce Department directive citing nationalist information concerns. Because verifying idiosyncratic nationality per API league successful existent clip is not technically feasible, Anthropic removed some models from the marketplace wholly alternatively than effort selective enforcement.

Semafor, successful an exclusive report by Reed Albergotti published this week, said the export power determination was partially motivated by suspicions that a China-linked radical had accessed Mythos. If the Chinese authorities obtained the model, U.S. officials feared Beijing could reverse-engineer it done a process known arsenic “distillation,” efficaciously replicating its capabilities.

It remains unclear which circumstantial enactment was involved, however entree occurred, oregon however the White House learned of it.

What Is Mythos

Anthropic launched Claude Mythos Preview successful April 2026 arsenic a frontier AI exemplary built astir autonomous cybersecurity discourtesy and defense. The institution deemed it excessively unsafe for nationalist release, citing its quality to place and exploit antecedently chartless bundle vulnerabilities astatine scale.

Key capabilities documented by Anthropic include:

• Autonomous recognition of thousands of high-severity zero-day vulnerabilities crossed large operating systems and browsers
• Discovery of a 27-year-old distant clang vulnerability successful OpenBSD, utilized successful captious infrastructure
• Chained Linux kernel exploits enabling afloat privilege escalation
• 83.1% accuracy connected the CyberGym vulnerability reproduction benchmark

Anthropic estimated AI-accelerated attacks could compress what presently takes months of adept enactment down to minutes. Cybercrime already costs an estimated $500 cardinal annually worldwide.

Access was restricted to astir 12 named firm partners and 40-plus further organizations done Project Glasswing, a antiaircraft inaugural designed to find and spot vulnerabilities earlier malicious actors could exploit them.

The Fable 5 Jailbreak

A 2nd model, Fable 5, a consumer-facing mentation of Mythos built with guardrails against precocious cyber use, was released successful aboriginal June. White House AI advisor David Sacks posted connected X that Anthropic CEO Dario Amodei was notified that Fable 5 had been jailbroken but reportedly downplayed the hazard and declined to code it.

Sacks said the medication “values Anthropic’s method capabilities” and described the concern arsenic “serious but easy resolved,” adding that “the shot is successful Anthropic’s court.”

Anthropic disputed the severity of the jailbreak, stating it was constrictive successful scope and progressive issues already contiguous successful different publically disposable models. The institution besides stated the White House did not rise Chinese entree concerns during nonstop discussions, which focused connected Fable 5 and the bypass technique.

Why Crypto Platforms Are Watching

Mythos has nonstop relevance to integer plus security. Anthropic’s enactment utilizing the exemplary identified millions successful imaginable blockchain exploit exposure, and the exemplary tin autonomously probe smart contracts, decentralized finance ( DeFi) protocols, APIs, and custody systems for weaknesses.

Major exchanges, including Coinbase, had reportedly entered discussions astir antiaircraft entree done Project Glasswing. The planetary entree suspension halts those arrangements for non-approved parties and removes a important antiaircraft instrumentality from operators extracurricular the approved spouse list. More recently, conscionable earlier Claude Fable 5 and Mythos 5 were banned, Zcash laminitis claimed Mythos audited the full Zcash concatenation and recovered nary bugs.

Security professionals who relied connected the model’s vulnerability scanning capabilities present person nary entree path. That spread benefits immoderate histrion that whitethorn person already obtained a copy.

US-China AI Competition

The incidental fits a broader pattern. Semafor’s reporting and anterior disclosures link Chinese state-linked actors to earlier attempts astatine utilizing Anthropic models for cyber espionage. A run documented by Anthropic, labeled GTG-1002, linked Chinese authorities actors to AI-assisted attacks against astir 30 entities including fiscal institutions.

Germany’s cybersecurity officials person separately warned that Chinese AI providers look to beryllium reducing nationalist updates portion gathering Mythos-equivalent capabilities internally. China has the information halfway infrastructure and compute capableness to prosecute that improvement path.

The suspension of Mythos and Fable 5 entree remains successful effect arsenic of June 14, 2026. Anthropic has not announced a timeline for restoring service.