How AI Changes the Speed of Public Exploits

This research post demonstrates that large language models (Claude) can rapidly produce functional exploits and comprehensive attack-surface mappings: the authors obtained root remote code execution against EPyT-Flow CVE-2026-25632 on first attempt, automated enumeration found 79 gadgets (including multiple RCE and SSRF paths), and the model was used to reason through and bypass ModSecurity/OWASP CRS rules, resulting in two responsibly disclosed WAF evasion findings. The report highlights how AI compresses manual research timelines from days to hours, lowering the barrier for attackers while offering defenders the same capabilities to triage and remediate at scale.