Anthropic Limits Claude Mythos Preview to Select Partners
Anthropic is not releasing Claude Mythos Preview to the general public. According to Anthropic, Mythos is the company's most capable model for cybersecurity work, and the risk of misuse is too high for a normal release.
What Anthropic Has Said About Mythos
Anthropic describes Mythos Preview as strong across many tasks and especially strong at computer security work. In the company's published materials on the model, it states that Mythos can identify and work with software vulnerabilities at a level higher than prior Claude models. Full attack claims and quantitative numbers are sourced from the Anthropic red team writeup and the UK AI Safety Institute evaluation, and readers should review those documents directly.
Project Glasswing
Access is handled through a new initiative called Project Glasswing. Anthropic is giving Mythos to a small set of vetted partners, with a focus on cybersecurity defense. Anthropic has said the goal is to learn how a Mythos-class model can be deployed safely at scale before any wider release.
Public Exposure and Evaluation
According to press reporting, the model first came to light through a CMS misconfiguration on March 26, 2026. Anthropic then officially previewed it as Mythos Preview on April 8, 2026. The UK AI Safety Institute published its own review of Mythos cyber capabilities.
Why It Matters
The decision to keep the most capable model restricted is a notable choice for a frontier lab. Most major labs have shipped their strongest models to customers with modest guardrails. Anthropic's Project Glasswing approach puts access control at the center of release policy and will likely draw questions about how other labs handle similar risks.
Discussion
Sign in to comment. Your account must be at least 1 day old.