Project Glasswing: AI Detects Ten Thousand Vulnerabilities in a Month

Anthropic has published the first results of its Project Glasswing, where the Claude Mythos Preview model detected over 10,000 critical vulnerabilities in 50 organizations within a month. The detection rate increased tenfold, with validations from Cloudflare, Mozilla, and Microsoft. The system autonomously scanned 1,000 open-source projects, identifying 23,019 flaws and generating functional exploits without human intervention.

autonomous AI security scanner interface, holographic display showing red vulnerability nodes spreading across network map, glowing exploit code fragments flowing from processor chip, multiple server racks in background with warning indicators flashing, digital shield being penetrated by luminous attack vectors, cinematic cyberpunk visualization, dark blue ambient lighting with orange hazard highlights, motion blur on data streams, ultra-detailed circuit board textures, photorealistic technical render

Claude Mythos: the first model to solve multistep attacks 🛡️

Claude Mythos Preview becomes the first AI model capable of solving multistep cyberattack simulations without human supervision. Its ability to autonomously build functional exploits marks a significant advancement in offensive cybersecurity. Findings from Microsoft and Mozilla confirm that the system not only identifies flaws but can chain them together to demonstrate real exploitability, surpassing traditional tools that require constant manual intervention.

The AI that hacks on its own: who watches the watchman? 🤖

Now we have an AI that finds vulnerabilities faster than an intern with coffee, and it also builds exploits while you think about updating your antivirus. Cloudflare and Mozilla are happy, but system administrators are already dreaming of nightly patches. Next thing you know, Claude will ask for a raise or demand root access as a working condition. Meanwhile, human hackers are wondering whether to ask the machine for a resume.