Anthropic Mythos Incident

Anthropic is investigating a report that unauthorized users accessed Claude Mythos Preview, its restricted cyber model, through a third-party vendor environment. (anthropic.com) (cybernews.com) Anthropic launched Project Glasswing on April 7, 2026, saying Claude Mythos Preview would be shared only with selected partners for defensive security work. The launch list included Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA and Palo Alto Networks. (anthropic.com) Anthropic said more than 40 additional organizations that build or maintain critical software infrastructure also received access, backed by up to $100 million in usage credits and $4 million in donations to open-source security groups. (anthropic.com) Mythos is a model built to find software flaws, the hidden coding mistakes that can let attackers break into systems. Anthropic said the same model is strong enough at finding and exploiting vulnerabilities that it would not release it to the public. (asisonline.org) (infosecurity-magazine.com) Reporting this week said a small group in a private Discord channel had been using Mythos after getting in through a contractor tied to Anthropic’s evaluation work. Anthropic said it had no indication the activity affected its own systems or extended beyond the vendor environment. (inc.com) (cybernews.com) That gap matters because Glasswing was presented as a controlled-access program built around the idea that only trusted defenders should get this kind of tool first. A leak through a vendor undercuts the central promise of that model: tight custody over a system designed to act like an automated vulnerability hunter. (anthropic.com) (cybernews.com) The Japan angle is less clear than the security incident itself. Publicly available reporting confirms Japanese companies and media have been tracking Glasswing closely, and Nikkei Asia reported on April 23 that NEC plans to tie up with Anthropic to embed Claude in products and systems, but I could not verify a separate official Japanese government program called “Project Glasswing.” (asia.nikkei.com) (japan.zdnet.com) What is verified is that Glasswing already spans U.S. tech groups, security vendors, financial firms and open-source infrastructure maintainers, and Anthropic has framed governments as necessary participants in the effort to secure critical software. (anthropic.com) (cyberscoop.com) The immediate next step is Anthropic’s investigation into how access was obtained and whether its vendor controls were enough for a model it chose not to release publicly. For now, the story is not that Mythos went wide; it is that a tool built for tightly limited access may have slipped outside that boundary almost as soon as it launched. (cybernews.com) (inc.com)