Krux
April 13, 2026
Anthropic Built an AI That Finds Zero-Days Autonomously
Published: April 13, 2026 at 12:30 AM
Updated: April 13, 2026 at 12:30 AM
100-word summary
Anthropic's new Claude Mythos Preview can scan operating systems and browsers to autonomously discover security vulnerabilities and write working exploits. The model found a 27-year-old bug in OpenBSD during testing. Only 40 partners get access (AWS, Apple, Google, Microsoft, CrowdStrike, Cisco), with up to $100M in credits to harden their software before attackers do. No public release is planned. Translation: AI just became useful enough at hacking that keeping it private feels safer than releasing it. The model costs $125 per million output tokens, meaning one breakthrough exploit could cost less than dinner.
What happened
Anthropic's new Claude Mythos Preview can scan operating systems and browsers to autonomously discover security vulnerabilities and write working exploits. The model found a 27-year-old bug in OpenBSD during testing. Only 40 partners get access (AWS, Apple, Google, Microsoft, CrowdStrike, Cisco), with up to $100M in credits to harden their software before attackers do. No public release is planned. Translation: AI just became useful enough at hacking that keeping it private feels safer than releasing it.
Why it matters
The model costs $125 per million output tokens, meaning one breakthrough exploit could cost less than dinner.