Researchers developed CVE-Genie, an automated multi-agent AI framework that converts CVE entries into reproducible exploits. The system orchestrates specialized agents to analyze vulnerability descriptions, rebuild the vulnerable environment, generate proof-of-concept code, and verify successful exploitation. Basically, it turns abstract security advisories into working attack code: 428 of 841 CVEs published in 2024–2025 (about 51%), at an average cost of $2.77 per CVE. For comparison, the best agent framework on CVE-Bench, a hand-built benchmark of critical web application CVEs from March 2025, managed to exploit at most 13% of its vulnerabilities. What could possibly go wrong with automating exploit development?
