Krux
April 8, 2026
Anthropic Built an AI That Finds Bugs Too Dangerous to Release
Published: April 8, 2026 at 12:39 AM
Updated: April 8, 2026 at 12:39 AM
100-word summary
Anthropic's new Claude Mythos Preview autonomously discovered critical vulnerabilities in OpenBSD, FFmpeg, and the Linux kernel. The model performed so well at finding security holes that the company won't release it publicly. Instead, Microsoft, Apple, Google, and 40 other partners get private access through Project Glasswing to hunt bugs before attackers do. Anthropic is backing the effort with $100 million in usage credits. On standard security tests, Mythos Preview jumped from 81% to 94% compared to Opus 4.6. The twist: building AI powerful enough to defend critical software means also building AI powerful enough to exploit it.
What happened
Anthropic's new Claude Mythos Preview autonomously discovered critical vulnerabilities in OpenBSD, FFmpeg, and the Linux kernel. The model performed so well at finding security holes that the company won't release it publicly. Instead, Microsoft, Apple, Google, and 40 other partners get private access through Project Glasswing to hunt bugs before attackers do. Anthropic is backing the effort with $100 million in usage credits. On standard security tests, Mythos Preview jumped from 81% to 94% compared to Opus 4.6.
Why it matters
The twist: building AI powerful enough to defend critical software means also building AI powerful enough to exploit it.