Summary
- Claude Mythos 5, a specialized AI model for cybersecurity and scientific research, has been launched by Anthropic.
- The company also introduced Claude Fable 5, a public-friendly version that directs sensitive inquiries to a less advanced model.
- Mythos has faced criticism for its ability to autonomously uncover vulnerabilities and simulate intricate cyberattacks.
On Tuesday, Anthropic officially unveiled Claude Mythos 5, an AI model designed for restricted access that boasts exceptional cybersecurity features, alongside the more accessible Claude Fable 5.
In a blog entry, Anthropic indicated that Mythos 5 will be available initially only to select cybersecurity entities, critical infrastructure operators, government collaborators, and certain life sciences researchers. The company warns that the Mythos-class AI models “have reached a threshold where they present significant risks.”
“We commenced Project Glasswing in April, releasing the inaugural Mythos-class model (Claude Mythos Preview) to a limited audience of cyber defenders and essential software infrastructure providers,” Anthropic stated. “During that time, we expressed our intention to eventually offer Mythos-level capabilities to all users, contingent on our development of robust safeguards to prevent misuse.”
This launch follows several months of scrutiny regarding Mythos' capacity to identify software vulnerabilities and execute complex cybersecurity tasks after information about the AI leaked online in March. Earlier this year, the U.K.'s AI Security Institute reported that a preview version was the first AI model to complete a 32-step corporate network intrusion exercise independently. In April, Mozilla disclosed that a preview of Mythos identified over 271 vulnerabilities in the Firefox browser.
Presenting Claude Fable 5: a Mythos-class model that has been made safe for general use.
Its capabilities surpass those of any model we've previously made available. pic.twitter.com/2AvmEjHIX8
— Claude (@claudeai) June 9, 2026
Until now, Anthropic has limited access to Mythos through its Project Glasswing initiative while researchers assessed the potential risks. In a subsequent post on X, Anthropic mentioned it would continue to broaden access to Mythos “through a wider trusted access program,” aimed at cybersecurity efforts and biomedical research.
The debut of Mythos coincides with a perceived thaw in relations between Anthropic and the administration of President Donald Trump, following a public disagreement regarding the application of Claude in military contexts.
“In collaboration with the U.S. government, we aim to gradually expand access to Claude Mythos 5, continuing our periodic inclusion of new partners, along with pursuing a trusted access program that allows cybersecurity firms to apply in a more systematic manner,” Anthropic noted, also indicating that users of Mythos Preview could upgrade to Mythos 5.
Recent reports suggested that Anthropic had embedded engineers within the National Security Agency to assist in deploying its Mythos AI model for cybersecurity initiatives, reportedly focusing on China and Iran.
Alongside Mythos 5, Anthropic introduced Claude Fable 5, a Mythos-class AI model that shares the foundational design of Mythos 5 but includes additional protections for general users.
“Fable 5’s capabilities outshine those of any model we’ve previously released to the public,” Anthropic stated. “It ranks at the forefront in nearly all evaluated AI capability benchmarks, demonstrating remarkable performance in software engineering, knowledge work, vision, scientific research, and more.”
According to Anthropic, both Mythos 5 and Fable 5 received top scores across various benchmarks in software engineering, cybersecurity, biology, and reasoning. The company reported a 78% score on ExploitBench, which assesses vulnerability detection and exploitation, and an 88% score on the Terminal-Bench coding benchmark. These results indicate that the models can manage longer and more intricate tasks compared to earlier Claude versions.
While Anthropic stated that Fable 5 excels at handling lengthy, complex tasks, it ensures that inquiries related to cybersecurity, biology, chemistry, or AI model replication are processed by Claude Opus 4.8—a safeguard that the company claims is activated in fewer than 5% of sessions.
“Launching a model with such capabilities entails risks,” Anthropic cautioned. “Without appropriate safeguards, Fable 5’s abilities in fields like cybersecurity could be exploited to inflict significant harm.”
This rollout occurs amid increasing scrutiny of how AI developers should introduce increasingly advanced systems. Anthropic maintains that Mythos could assist defenders in identifying and rectifying vulnerabilities before they can be exploited by adversaries. Conversely, critics, including OpenAI CEO Sam Altman, have accused the company of employing “fear-based marketing” to promote its AI models.