Anthropic Reveals Claude Mythos Preview: AI That Can Hack Every Major OS Is Too Dangerous for Public Release

Anthropic announced Claude Mythos Preview—its most capable AI model—while simultaneously declaring it will not be publicly released due to unprecedented cybersecurity risks. In testing, Mythos autonomously identified thousands of high-severity vulnerabilities across every major operating system and browser, including a 17-year-old remote code execution vulnerability in FreeBSD that grants root access from the internet. The model scored 93.9% on SWE-bench Verified and 94.6% on GPQA Diamond—benchmarks where top models typically score 50-70%. Instead of public release, Anthropic is deploying Mythos exclusively through Project Glasswing, a consortium of 40 partner organizations (including Google, Apple, Microsoft, Nvidia, Amazon, CrowdStrike) focused on defensive cybersecurity. The approach is unprecedented: a model too capable to release publicly but too important not to deploy at all. Pricing for partner access: $25 per million input tokens—5x the rate of Claude Opus 4.6.