Cryptographer: AI labs should track whether AI security tools actually catch real vulnerabilities
matthew_d_green · x · 2026-10-22
Cryptographer Matthew Green poses a key measurement question: for every major vulnerability found in the coming months, did maintainers run AI security tools like Mythos or Daybreak, or nothing at all? Someone should track this data to count how many relevant hardening fixes went unlanded. He adds that labs should have mandated this data collection as a condition of researcher access programs — but evidently nobody is keeping it.
More from Safety
- Ed Simeoni: cybersecurity needs 'cyber defense factories' running frontier models 24/7 — AccBalanced · 2026-10-22
- Econoar on AI Hacking: Short-Term Attackers Win, Long Term It's a 'Hardening Era' — econoar · 2026-10-22
- Employees fired over false AI-detector positives are suing — and winning settlements — DavidLinthicum · 2026-10-22
- Security researcher: today's CVE flood reflects better discovery tools, not new flaws — realsohamparekh · 2026-10-22
- HEIF Heist: image parser bugs chained into RCE on OpenAI, Meta, Slack and more — moyix · 2026-10-22
- Xbow to host Reddit AMA with security researcher moyix on Oct 22 — moyix · 2026-10-22