Anthropic’s safety warnings may have just backfired — the government has pulled the plug on its most powerful AI

Details

On Friday at 5:21 PM ET, Anthropic received a government directive to shut off access to both Claude Fable 5 and Claude Mythos 5 for all users worldwide — not just the foreign nationals that the export control order nominally targeted. Anthropic complied but published a detailed blog post disagreeing with the action. Fable 5, released just three days earlier, had immediately ranked as the most capable publicly available model per Vals AI benchmarks. Mythos is Anthropic's most capable model, previewed in April and kept restricted ever since because of its ability to identify security vulnerabilities in every major OS and browser tested. Anthropic released it only through Project Glasswing, a controlled program shared with ~50 vetted organizations including Amazon, Apple, Google, Microsoft, and CrowdStrike. Fable 5 was the commercial version — same power, but with guardrails blocking high-risk domains like cybersecurity and biology. The government's stated concern is a claimed jailbreak of Fable 5. Anthropic says it received only verbal evidence of a 'narrow, non-universal' jailbreak that amounts to prompting the model to read a codebase and identify flaws — a capability already present in GPT-5.5 and routinely used by cybersecurity professionals. Anthropic argues its actual safety infrastructure consists of independent classifier systems that operate separately from the model itself, meaning a successful prompt jailbreak does not bypass the underlying protections against truly harmful outputs. This is the first known instance of the U.S. government ordering a commercial AI model taken offline globally over a security concern, setting a major precedent for how frontier AI capabilities may be regulated going forward.