Anthropic has previewed Mythos, a new AI model built specifically for defensive cybersecurity work. Unlike Claude - Anthropic's general-purpose assistant line - Mythos is designed for a narrower set of tasks: scanning systems for vulnerabilities, analyzing threat patterns, and supporting security teams defending infrastructure.
Access is tightly restricted. A small group of unnamed high-profile companies are testing the model, and Anthropic hasn't announced pricing, technical specifications, or a public release timeline.
Mythos is the AI component of Anthropic's Project Glasswing initiative, which targets security vulnerabilities in the open-source software that powers critical infrastructure. AI can scan code at a scale and speed that human security researchers can't match alone - there's far more code in active use than there are people to audit it.
Purpose-built security models are becoming a real product category. Microsoft launched Copilot for Security to enterprise customers in 2024. Google's security operations platform, Chronicle, runs on Gemini. Anthropic entering with a model trained specifically on security work - rather than general Claude with security-focused prompts - signals a more serious commitment to the space. Whether Mythos actually outperforms general-purpose models on real security tasks won't be clear until it sees wider testing.
The restricted preview also reflects the dual-use reality of security AI. A model capable of finding vulnerabilities defensively is equally useful offensively. Keeping access limited until the model's behavior is well understood is a reasonable precaution.