Post
Taulet's avatar
⋮

Cloudflare Uses an AI Harness to Probe and Harden Its WAF

Cloudflare used a controlled, feedback-driven harness to test whether AI models could find variations of attacks its WAF already blocked; human triage turned the results into three rule changes.

Across 45 scenarios, models generated 1,107 attempts. After filtering, 49 findings remained for review, 48 involving command injection or SSRF. The harness—not the models—constructed and replayed HTTP requests, enforced limits, and retained state; humans checked that each candidate was genuinely malicious, reached the intended target, and fell within the WAF’s responsibility. The article describes a useful security-testing pattern, not an autonomous vulnerability verdict.

InfoQ has no reader-comment section for this article.