Anthropic launches restricted Mythos 5, releases safer Fable 5

On June 9, 2026, Anthropic released Claude Mythos 5 as a restricted-access AI model for approved cybersecurity firms, critical infrastructure operators, government partners and selected life-sciences researchers. At the same time the company introduced Claude Fable 5, a general-use variant that includes additional safeguards.

Mythos 5 follows an April pilot called Project Glasswing that gave a preview of the model to a small group of cyber defenders and infrastructure providers. A preview completed a 32-step corporate network intrusion exercise without human assistance and identified more than 270 security defects in a major web browser.

Anthropic wrote that Mythos-class models had “reached a threshold where they present significant risks.” The company plans to limit Mythos access initially through an expanded trusted-access program and to add partners periodically.

The company reported that Mythos 5 and Fable 5 scored highly on internal tests, including 78% on ExploitBench, which measures vulnerability discovery and exploitation, and 88% on a terminal coding benchmark. Anthropic described the models as able to handle longer, more complex tasks than earlier Claude systems and noted Mythos 5 is intended for technical work such as vulnerability discovery and complex cyber exercises.

Fable 5 uses the same core architecture as Mythos 5 but routes requests that touch on cybersecurity, biology, chemistry or attempts to replicate AI models to a less capable system, Claude Opus 4.8. Anthropic wrote that the fail-safe is invoked in fewer than 5% of sessions and that Fable 5’s capabilities “exceed those of any model we’ve ever made generally available.”

The company plans to expand Mythos access “in consultation with the U.S. government” and to allow vetted organizations to apply through a trusted-access program. Users of the earlier Mythos preview can upgrade to Mythos 5. Some reports indicate Anthropic embedded engineers with a U.S. intelligence agency to help deploy the model for cybersecurity work; Anthropic did not provide operational details in its announcement.

Reactions to the launch varied. Some industry figures warned that promoting defensive uses while highlighting offensive capabilities could encourage misuse. A prominent technology executive described the approach as “fear-based marketing.” Anthropic acknowledged risks and wrote it will keep safeguards in place during the roll-out.

Researchers are increasingly using advanced AI models for code generation, vulnerability identification and simulated cyber operations. Companies are testing different access controls for high-capability models while continuing to develop safeguards.