Cloudflare’s AI-Driven WAF Testing Produces 49 Findings and Stronger SSRF Rules

Cloudflare used frontier AI models in a controlled, black-box testing harness to probe its Web Application Firewall (WAF), generating 1,107 attempts across 45 scenarios and retaining 49 findings after human triage. Models mutated previously blocked attack payloads using response feedback, while a Python harness executed requests, maintained state, enforced limits, and collected results. Of the retained findings, 48 involved command injection or server-side request forgery (SSRF). The exercise contributed to two new SSRF detections and an improvement to an existing rule in Cloudflare’s Managed Ruleset. Reviewers checked whether requests reached their targets, remained malicious, were clearly unblocked, and could be safely reproduced; redirects alone were not considered proof of exploitation.

The WAF exercise complements Cloudflare’s broader model-agnostic vulnerability research pipeline, which separates discovery from validation and uses different models for adversarial cross-checking. Cloudflare reported 7,245 actionable findings delivered to engineering teams across a validation workload spanning 145 repositories, with persistence, deduplication, cross-repository dependency tracing, proof-of-concept testing against untouched code, and mandatory human review before fixes are merged. Those code-audit findings came from an isolated research environment, not active unpatched production issues. Together, the efforts highlight a practical approach for security teams: use AI to expand testing coverage, but retain deterministic execution, independent validation, and human approval before treating results as confirmed vulnerabilities or deploying fixes.