Anthropic debuts preview of powerful new AI model Mythos in new cybersecurity initiative

Preview and partners
Anthropic on Tuesday released a preview of Mythos, a new “frontier” model it says will be used by a small coterie of partners for cybersecurity work. The limited rollout is part of Project Glasswing, a program in which more than 40 organizations — including Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft and Palo Alto Networks — will deploy the model to scan first‑party and open‑source software for vulnerabilities. Anthropic says the preview won’t be generally available; access is tightly controlled.
Capabilities, claims and leaks
Mythos is billed as a general‑purpose frontier model for Anthropic’s Claude systems with strong agentic coding and reasoning skills. Anthropic claims Mythos identified “thousands of zero‑day vulnerabilities, many of them critical”; it has been reported that many of those bugs date back one to two decades. A leaked draft blog — allegedly left exposed in a public data cache under the name “Capybara” — described the model as larger and more intelligent than prior Opus models and “by far the most powerful AI model we’ve ever developed.” That raises an obvious question: a tool built to find bugs can also help exploit them. Who watches the watchmen?
Context and the wider stakes
Project Glasswing is pitched as a public‑good play: partners will share findings so the broader industry benefits. But the timing is fraught. Anthropic says it has engaged in “ongoing discussions” with federal officials about Mythos; it has been reported that those talks overlap with a legal fight after the Pentagon labeled the company a supply‑chain risk. Add to that last month’s accidental exposure of nearly 2,000 source files and the takedown ripple on GitHub, and you get a clear emotional chord — excitement for a potent new security tool, mixed with uneasy distrust. It’s the classic dual‑use dilemma all over again: powerful tech, big promise, real risk.
Sources: techcrunch
Comments