Mythos Hits Exploits First-Try 83% of the Time — Too Hot to Ship

Anthropic's Mythos Preview is being withheld from public release because its offensive cyber capabilities are simply too potent. In testing, the model successfully reproduced vulnerabilities and generated working proof-of-concept exploits on the first attempt in 83.1% of cases — entirely autonomously, with no human steering.

The model swept through every major OS and browser, surfacing thousands of previously unknown software flaws, including a 27-year-old OpenBSD remote crash bug and chained Linux kernel exploits that grant full machine control. Anthropic says Mythos is "currently far ahead of any other AI model in cyber capabilities."

Rather than a public launch, Anthropic is rolling out a controlled early-access program called Project Glasswing — giving defenders a head start to patch systems before the wider AI community gets similar tools. The bet: offense is already coming, so arm the defense first.

Why it matters: when an AI can find and exploit decades-old bugs faster than entire red teams, the security industry's timelines just collapsed.