Anthropic Debuts Mythos, Its Most Capable Model Yet, Restricted to 12 Cybersecurity Partners Under Project Glasswing

On April 7, 2026, Anthropic unveiled a preview of Mythos — described internally as the company's most capable model to date — but deliberately restricted access to a curated group of 12 partners for defensive cybersecurity work under a program called Project Glasswing. Partners include Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks. Anthropic says Mythos far exceeds its public models on coding, academic reasoning, and cybersecurity benchmarks.

The restricted rollout is a direct result of safety concerns: Anthropic withheld a broader release specifically because of the risk that Mythos could be weaponized to find and exploit software vulnerabilities at scale. A follow-up TechCrunch analysis published April 9 explored whether the limited rollout was driven primarily by safety concerns for the broader internet or by competitive and reputational considerations for Anthropic — a question the company did not directly answer. The decision reflects Anthropic's long-standing "responsible scaling policy" which ties deployment breadth to safety evaluation outcomes.

Project Glasswing marks a notable inflection point in how frontier AI labs handle their most capable models. Rather than the typical phased public rollout, Anthropic is using a curated partner model — effectively a closed beta with institutional accountability — to gather real-world safety data before any public release. Industry observers note the approach resembles how biosafety labs handle dual-use research: controlled access for verified defensive purposes, with no timeline given for broader availability. Anthropic has not confirmed when or whether a public version of Mythos will be released.