Anthropic's Claude Mythos Preview autonomously discovered thousands of zero-day vulnerabilities across major operating systems, browsers, and critical software including OpenBSD, Linux kernel, and FFmpeg. The model developed exploits entirely without human steering, including a 27-year-old OpenBSD privilege escalation and a 16-year-old FFmpeg flaw that evaded automated testing. This demonstrates significant advancement in AI-powered vulnerability research with major implications for infrastructure security.
Safety
Project Glasswing: Securing critical software for the AI era
Claude Mythos autonomously discovered thousands of zero-day vulnerabilities—including exploits for a 27-year-old OpenBSD flaw and 16-year-old FFmpeg bug—marking a significant leap in AI-powered vulnerability research and infrastructure security.
Tuesday, April 7, 2026 12:00 PM UTC2 MIN READSOURCE: Hacker NewsBY sys://pipeline
Tags
safety
/// RELATED
Infrastructure6d ago
Soft launch of open-source code platform for government
Dutch government launches code.overheid.nl on Forgejo to achieve digital sovereignty and replace commercial GitHub dependency across all government bodies.
Products6d ago
Zed 1.0
Zed code editor reaches 1.0 after five years with custom GPU-accelerated rendering (GPUI) and launches DeltaDB, a CRDT engine enabling real-time human-AI code collaboration.