Claude Mythos Preview & Project Glasswing: what Anthropic publishes—and why release is constrained
You may hear buzz about Claude Mythos Preview online. This article sticks to Anthropic.com and Anthropic's linked Frontier Red Team materials—so you can verify claims in primary documents rather than social rewrites.
Project Glasswing (official page)
Anthropic's Project Glasswing page describes a coalition of companies working on securing critical software using frontier model capabilities, anchored around Claude Mythos Preview—described there as a general-purpose, unreleased frontier model trained by Anthropic. The page argues that such models can find serious vulnerabilities at scale, and that defensive collaboration is urgent as capabilities spread.
Read the full narrative and partner statements at anthropic.com/glasswing.
The same initiative page includes a quantitative example from Anthropic's evaluation context: on CyberGym "Cybersecurity Vulnerability Reproduction", it reports Mythos Preview 83.1% vs Claude Opus 4.6 66.6%—again, see the Glasswing page for the exact table and footnotes.
Why there is no broad "public Mythos API" (per Anthropic)
Anthropic's public writing ties limited release to misuse risk at very high capability on cybersecurity tasks, and to the goal of learning safeguards on less capable models first. The Claude Opus 4.7 announcement explicitly states that Opus 4.7 is the first model shipping with automated blocks for prohibited/high-risk cybersecurity uses, and that lessons from deployment should help toward eventual broader release of Mythos-class models—see anthropic.com/news/claude-opus-4-7.
In other words: Anthropic's stated strategy is not "hide the model forever," but constrain access, pair it with defensive programmes, publish technical details where responsible, and iterate on safety systems before general availability.
Primary documents to read next
For students and builders
Policy and safety discussions move quickly. For coursework, cite the exact PDF/HTML you read (Anthropic post + date), separate vendor claims from independent audits, and pair frontier-model news with fundamentals: evaluation basics, agentic AI, and Opus 4.7.