Anthropic, the AI safety-focused startup behind the Claude models, has quietly unveiled a preview of its most advanced system yet. Named Mythos, it forms the centrepiece of Project Glasswing, a cybersecurity collaboration with 12 heavyweight partners.
These include Amazon, Apple, Cisco, CrowdStrike, Microsoft, and Palo Alto Networks. The group will deploy the model for defensive operations and to fortify critical software.
Though not specifically engineered for security tasks, Mythos leverages its formidable agentic coding and reasoning strengths. It sifts through proprietary and open-source codebases for vulnerabilities. Anthropic says the model has already detected thousands of zero-day flaws in recent weeks. Many are critical, and some have lingered undetected for up to two decades.
A leaked internal memo earlier this year described it as one of the company’s “most powerful” creations. Leaked details revealed it as “by far the most powerful AI model we’ve ever developed.” This marks a step beyond its Opus predecessors in areas like software coding, academic reasoning and cybersecurity.
The partners will exchange insights from their trials to benefit the wider industry. Access remains restricted to these 12 firms plus 40 others. This limited rollout underscores Anthropic’s cautious approach amid growing scrutiny over powerful AI.
Complicating matters are ongoing talks with U.S. federal officials about Mythos’s potential uses. These sit against a heated legal feud with the Trump administration. The Pentagon has flagged Anthropic as a defence supply-chain risk. This stems from the firm’s refusal to enable AI for autonomous targeting or surveillance of Americans.
Mythos’s existence first leaked last month. A draft post codenamed Capybara appeared in an unsecured public data lake. Security researchers caught the blunder, which Anthropic blamed on human error. The firm has since weathered more embarrassment. An accidental dump of thousands of code files came during a Claude Code update. A botched cleanup then yanked unrelated GitHub repositories offline.
Project Glasswing positions Mythos as a force for good in cybersecurity. Yet it also highlights the tightrope AI developers walk between innovation and risk.