We tested our own WAF with frontier AI models. Here’s what we found

Cloudflare evaluated its web application firewall using an AI-driven dynamic testing system that simulated an attacker by iteratively mutating known payloads across 45 scenarios involving XSS, SQL injection, command injection, SSRF, path traversal, Log4j, and log injection in an authorized customer staging environment. The models had no access to source code, WAF rules, or enforcement controls, while Python-based safeguards controlled request replay, allowlisted targets, limits, and evidence collection. Of 1,107 generated attempts, 558 were blocked, and human triage reduced unblocked or ambiguous outputs to 49 WAF-relevant findings, mostly involving command injection and SSRF; these findings were leads for validation rather than evidence of successful exploitation. One SSRF example showed that a trailing-dot representation of a cloud metadata IP address produced a redirect instead of a WAF block, prompting investigation without proving origin access. The exercise led to new detections for obfuscated SSRF hosts and restricted protocols, plus improvements to an existing SSRF rule, and indicated that broader scenario coverage was more useful than simply increasing attempts within a single scenario. Cloudflare recommends defense in depth, including properly configured managed WAF rules and attack scoring, staged testing, positive security controls, additional API and bot protections, and regular software patching because WAF bypasses still require an exploitable application to cause harm.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.