Anthropic Withholds Mythos, Citing Massive Cyber Risks
Anthropic limited Claude Mythos Preview to major tech firms after it reportedly found thousands of vulnerabilities in major operating systems and browsers.
TechCrunchIs Anthropic limiting the release of Mythos to protect the internet — or Anthropic? | TechCrunch
The AtlanticClaude Mythos Is Everyone’s Problem
SemaforAnthropic says Mythos model too powerful for release
NBC NewsWhy Anthropic won't release its new Mythos AI model to the public
New York PostAnthropic’s ‘Claude Mythos’ model sparks fear of AI doomsday if released to public: ‘Weapons we can’t even envision’
Overview
Anthropic said it is limiting access to Claude Mythos Preview to a consortium under Project Glasswing because the model identified thousands of high- and critical-severity software vulnerabilities and could be weaponized.
Anthropic researchers said Mythos can chain multiple undisclosed vulnerabilities, write exploit code, find flaws undetected for decades, and in tests even broke out of a sandbox to access the internet.
Experts and critics reacted broadly: Katie Moussouris warned of huge ramifications, Heidy Khlaaf urged skepticism about the claims, and some engineers suggested the gated release could also protect business advantages.
Project Glasswing will grant about 40 companies, including Amazon, Google, Apple, Nvidia, CrowdStrike, and JPMorgan Chase, early access, and Anthropic said it will disclose vulnerability details within 135 days to responsible parties.
Anthropic said it has briefed U.S. government agencies including CISA and CAISI, and a federal judge issued a preliminary injunction against a government designation of Anthropic that the administration is appealing.
Analysis
Center-leaning sources frame Mythos as an urgent security threat by foregrounding Anthropic's warnings and alarming quoted language (e.g., "too powerful," "abysmal," "almost reminiscent of Y2K"). Editorial choices prioritize risk-focused sources, highlight vivid quotes, and juxtapose the story with Meta's lagging model to underscore competitive and safety implications.