Anthropic just made a calculated bet: the safest way to deploy its most cyber-capable model is to never let you talk to it. Claude Security scans now run on Mythos 5, the model class that until yesterday reached only vetted defenders through Project Glasswing. Enterprise customers can connect a GitHub repository and get back a fixed artifact — CWE category, confidence and severity ratings, a suggested patch — with no prompt box in sight. The model traces data flows across files and reads Git history rather than pattern-matching against rules, targeting memory corruption, injection flaws, authentication bypasses, and cross-file logic errors. Anthropic claims each finding passes an adversarial verification step where the model challenges its own result before surfacing it, which is their stated mechanism for cutting false positives.
There’s no separate model add-on. Mythos 5 scans bill as standard token usage under the existing Enterprise plan, enabled by an admin in the console. But the deployment path is narrow. Startups and mid-market teams on Pro, Max, or Team plans don’t get the Mythos-powered scan at all. The Claude Security plugin for Claude Code is in beta for all users, but it runs on whatever models are already in your account — not Mythos 5. Patching happens separately too: users open Claude Code on the web to implement fixes, and that step uses the organization’s existing models. Every patch requires human review and approval.
The packaging is the entire point. Anthropic has said Mythos 5 holds the strongest cybersecurity capabilities of any model in the world, and the same skills that find vulnerabilities can write exploits. Claude Fable 5 — the same underlying model shipped broadly — routes cyber, biology, and chemistry queries away through classifiers. This release extends that reasoning: risk concentrates when a user can prompt the model directly. A scan result, by contrast, is a defined output generated by Mythos but never steerable by the person receiving it. The same logic applies to partner integrations, where end users get a patch list or alert without any path to ask for an exploit.
Two other items landed in the same announcement. Anthropic launched the Defender Advantage Fund (0xDAF), offering $35 million in Claude credits to organizations securing open-source software. And the Cyber Verification Program will expand over the coming weeks to cover broader dual-use capabilities on Opus and Sonnet, with Mythos-class access to follow. For security teams drowning in SAST false positives, the verification step is the thing to watch — if it works, this could reset expectations for what automated code review can do. If it doesn’t, it’s just another scanner with better marketing.
