Anthropic's leaked $80 Claude Mythos model is a terrifyingly good zero-day hacker

Red team testing for the upcoming model, codenamed Oceanus, came to a grinding halt after Chinese API aggregators began reselling unauthorized access to the system. This unsanctioned preview exposed the eye-watering price tag of 16 dollars per million input tokens and 80 dollars per million output tokens, making it a luxury tool for the cyber-elite.

The model, expected to launch publicly as Claude Mythos, features a massive one-million-token context window, matching the current high-end standards. But instead of just summarizing endless corporate PDFs, this AI is designed to hunt for zero-day vulnerabilities and write functional exploits completely on its own. It is basically a digital hacker that does not drink energy drinks or complain about working weekends.

During its secretive closed testing phase under Project Glasswing, which expanded from fifty organizations to two hundred critical infrastructure sites, the AI achieved a staggering 83% success rate in generating working exploits on its first attempt. It single-handedly discovered over ten thousand high and critical vulnerabilities across every major operating system and web browser.

Among its most embarrassing trophies are a twenty-seven-year-old vulnerability in OpenBSD and a sixteen-year-old bug in FFmpeg, both of which had quietly survived decades of human security audits. If humans cannot find a bug in a quarter of a century, perhaps it is time to let our new silicon overlords audit the code.

Releasing a tool this powerful is the ultimate tech-industry paradox. A piece of software designed to patch the world's crumbling digital infrastructure is simultaneously the most potent, automated weapon ever handed to malicious actors, leaving global cybersecurity hanging by a thread.

Source: X