Claude Mythos Preview Scores 73% on Elite Cybersecurity Tests

Anthropic's Claude Mythos Preview just put up serious numbers in cybersecurity. The AI Security Institute ran evaluations on the model — announced April 7th — and found it achieved a 73% success rate on expert-level capture-the-flag challenges.

Here's the kicker: no AI model could complete these challenges before April 2025. That's a meaningful leap in capability for automated cyber offense and defense.

The AISI specifically designed the evaluation to assess cybersecurity capabilities, putting Claude Mythos Preview through the kind of real-world hacking scenarios that separate toy demos from genuine threats.

A 73% success rate on problems that were previously unsolvable by any model is a significant benchmark. It raises big questions about how quickly AI-powered cybersecurity tools — and AI-powered attacks — are advancing.