NEWS · MODELS · #1658
Cloudflare uses AI harness to probe and harden its WAF
Cloudflare placed frontier AI models inside a controlled Python harness to generate and mutate attack payloads against its Web Application Firewall (WAF). Across 45 scenarios the system produced 1,107 mutation attempts, leaving 49 post-triage findings; the experiment contributed to three changes in Cloudflare’s Managed Ruleset (two new SSRF detections and an improvement to an existing SSRF rule), with human reviewers validating results before rule changes.
KEY POINTS
- Cloudflare placed frontier AI models inside a controlled Python harness to generate and mutate attack payloads against its Web Application Firewall (WAF).
- Across 45 scenarios the system produced 1,107 mutation attempts, leaving 49 post-triage findings; the experiment contributed to three changes in Cloudflare’s Managed Ruleset (two new SSRF detections and an improvement to an existing SSRF rule), with human reviewers validating results before rule changes.
- This shows practical use of frontier models in a constrained discovery-and-validation harness to find exploitable variations and produce actionable WAF rule updates while retaining human review.
WHY IT MATTERS
This shows practical use of frontier models in a constrained discovery-and-validation harness to find exploitable variations and produce actionable WAF rule updates while retaining human review.