AI-written patches become attack maps for hackers
The vm2 library patched 34 sandbox-escape flaws, 19 co-written by Anthropic's Claude. Researchers say AI agents can analyze public patches to find bypasses, turning fixes into blueprints for attackers.