Anthropic launched Claude Mythos Preview — a restricted-access model that autonomously identifies critical security flaws across major platforms, available only to vetted enterprise partners.
Anthropic announced Project Glasswing, a cybersecurity initiative partnering with Nvidia, Google, AWS, Apple, and Microsoft. The program introduces Claude Mythos Preview, a general-purpose model not planned for public release due to security concerns. The model autonomously identified vulnerabilities in every major OS and web browser. Access is restricted to prevent adversarial exploitation, and Anthropic has briefed the government but declined to share specifics.
Claude Mythos Preview represents a step-change in agentic security tooling — it found critical vulnerabilities in every major OS and browser through autonomous reasoning, not custom security training. This isn't a SAST tool or a code linter; it's reasoning-based, agentic vulnerability discovery at system scope. The architecture signals that general-purpose reasoning models with strong coding ability are crossing into red-team territory that was previously human-expert-only.
If you own security posture for a product, run your current static analysis toolchain against your last 3 CVEs and ask Claude Sonnet to reason through each one — benchmark whether it surfaces the root cause faster than your existing tools.
Go to claude.ai and open a new conversation
Tags
Related
Also today
Signals by role
Also today
Tools mentioned