Krux
April 11, 2026
Anthropic Built an AI That Finds Zero-Days, Won't Release It
Published: April 11, 2026 at 12:35 AM
Updated: April 11, 2026 at 12:35 AM
100-word summary
Anthropic just announced Claude Mythos Preview, a model that can autonomously discover and exploit zero-day vulnerabilities in software. The catch? You can't use it. Instead, Anthropic is giving restricted access to AWS, Apple, Microsoft, CrowdStrike, and 40 other partners through Project Glasswing to hunt bugs in critical infrastructure before attackers do. Partners get $100M in usage credits to scan everything from operating systems to browsers, with findings published in a 90-day public report. The model stays locked down because the same AI that patches your software could wreck it. Pricing after preview: $25 per million input tokens. Translation: the race between AI-powered attackers and defenders just became asymmetric, and Anthropic...
What happened
Anthropic just announced Claude Mythos Preview, a model that can autonomously discover and exploit zero-day vulnerabilities in software. The catch? You can't use it. Instead, Anthropic is giving restricted access to AWS, Apple, Microsoft, CrowdStrike, and 40 other partners through Project Glasswing to hunt bugs in critical infrastructure before attackers do. Partners get $100M in usage credits to scan everything from operating systems to browsers, with findings published in a 90-day public report. The model stays locked down because the same AI that patches your software could wreck it. Pricing after preview: $25 per million input tokens.
Why it matters
Translation: the race between AI-powered attackers and defenders just became asymmetric, and Anthropic is betting controlled access beats an arms race.