Anthropic built an AI model so good at hacking that the company decided most people shouldn’t be allowed to use it.
On April 7, 2026, Anthropic announced the limited release of Claude Mythos Preview, its latest frontier model. Rather than making it broadly available, the company restricted access through something called Project Glasswing, a controlled program offering the model only to select tech firms and key infrastructure providers for defensive cybersecurity purposes.
What Mythos actually did
During internal testing, Mythos autonomously detected thousands of high-severity vulnerabilities across various software platforms. That number includes 271 vulnerabilities found in Firefox alone. Some of these bugs had been sitting dormant in legacy systems for decades, invisible to human security teams and prior AI models alike.
Mythos didn’t just find vulnerabilities. It demonstrated the ability to chain exploits together in sophisticated sequences, essentially showing how a bad actor could turn individual weaknesses into coordinated cyberattacks.
