Anthropic Unleashes Claude Mythos 5 for enterpriseAI, Plus the Secure and Streamlined Fable 5 for Everyone

On Tuesday, Anthropic unveiled Claude Mythos 5—a tightly controlled AI system it claims possesses groundbreaking cybersecurity abilities—alongside Claude Fable 5, designed for everyday users.

According to a company blog post, Mythos 5 will initially be accessible only to vetted cybersecurity firms, critical infrastructure operators, government collaborators, and select life sciences researchers. Anthropic stated that Mythos-class models “have reached a level where they pose substantial risks.”

“Back in April, we initiated Project Glasswing, granting early access to the first Mythos-class model (Claude Mythos Preview) exclusively to a small circle of cyber defense teams and critical software providers,” the company explained. “At the time, we expressed our goal of eventually making Mythos-level capabilities available to all users—provided we could develop robust enough safeguards to reliably prevent abuse.”

This rollout follows months of unease surrounding Mythos’ capacity to detect software weaknesses and carry out advanced cyber operations, especially after details about the AI surfaced online in March. Earlier this year, the U.K.’s AI Security Institute confirmed that a preview version became the first AI to autonomously complete a 32-step corporate network breach exercise. Separately, Mozilla reported in April that a Mythos preview identified more than 271 security flaws in the Firefox browser.

Until now, Anthropic had limited Mythos access through its Project Glasswing initiative while experts assessed potential dangers. In a subsequent post on X, the company announced plans to gradually broaden Mythos availability “via an expanded trusted access program” focused on cybersecurity and biomedical research.

The Mythos launch coincides with signs of improving relations between Anthropic and the Trump administration, following a public dispute over the military use of Claude.

“Working closely with the U.S. government, we intend to steadily widen access to Claude Mythos 5—regularly onboarding new partners and establishing a formal trusted access program so cybersecurity organizations can apply systematically,” Anthropic noted, adding that current Mythos Preview users can upgrade to Mythos 5.

Recent press reports indicate that Anthropic has stationed engineers at the National Security Agency to assist in deploying its Mythos AI for cyber operations, with reported focus on threats from China and Iran.

Alongside Mythos 5, Anthropic rolled out Claude Fable 5—a Mythos-class AI built on the same core architecture as Mythos 5 but enhanced with extra safety measures for public use.

“Fable 5 outperforms any model we’ve previously released to the general public,” the company stated. “It leads on nearly all standard AI capability benchmarks, excelling in software engineering, knowledge tasks, visual analysis, scientific inquiry, and numerous other domains.”

Anthropic reported that both Mythos 5 and Fable 5 achieved top rankings across multiple benchmarks in software engineering, cybersecurity, biology, and logical reasoning. Specifically, the models scored 78% on ExploitBench—a test measuring vulnerability detection and exploitation—and 88% on the Terminal-Bench coding challenge. The company emphasized that these results demonstrate the models’ ability to manage longer, more intricate tasks than earlier Claude versions.

While Fable 5 shines in handling extended, complex assignments, Anthropic has implemented a safeguard: requests related to cybersecurity, biology, chemistry, or AI self-replication are automatically routed to Claude Opus 4.8. The company noted this protective measure activates in fewer than 5% of user sessions.

“Deploying a model this powerful carries inherent risks,” Anthropic cautioned. “Without proper controls, Fable 5’s cybersecurity skills could be weaponized to inflict serious harm.”

The release unfolds amid intensifying debate among AI developers about responsibly launching ever-more-capable systems. Anthropic maintains that Mythos can empower defenders to patch vulnerabilities before malicious actors exploit them. However, critics—including OpenAI CEO Sam Altman—have accused the company of leveraging “fear-based marketing” to promote its AI offerings.