Anthropic announces Claude Mythos Preview and Project Glasswing

Background

Claude Mythos Preview was an unreleased frontier model. On 7 April 2026 Anthropic announced Project Glasswing, a programme that gave a restricted group access to the model so that it could be used to secure critical software. ^[1] Wikipedia notes that details of the model had leaked from a public database in late March 2026. ^[2]

What happened

Anthropic said Mythos Preview had found thousands of high-severity vulnerabilities, including in every major operating system and web browser. Most of them were found autonomously, meaning the model worked without step-by-step direction from a person. ^[1]

The partners named in the programme were AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks, along with more than 40 other organisations. ^[1]

Key details

Anthropic committed up to $100 million in usage credits and $4 million in donations to open-source security organisations. ^[1]

On benchmarks, Mythos Preview scored 93.9% on SWE-bench Verified and 83.1% on CyberGym. Claude Opus 4.6 scored 80.8% and 66.6% on the same two tests. ^[1]

Anthropic said it did not plan general availability until safeguards could block the model's most dangerous outputs. ^[1]

A vulnerability is a weakness in software that an attacker could exploit, and a high-severity vulnerability is one whose misuse would cause serious harm. Securing critical software means finding such weaknesses and fixing them before they can be used. A model that finds vulnerabilities autonomously can work through large codebases without a person directing each search, which is why Anthropic described its role in the programme in these terms.

Impact

Under Project Glasswing, access to Mythos Preview was limited to a restricted group, and the model was not released to the general public. ^[1]