We tested our own WAF with frontier AI models. Here’s what we found
Cloudflare used frontier LLMs as adaptive attackers to test its WAF, logging 1,107 attempts and hardening detections from bypasses.
Cloudflare built a Python-based WAF tester where an LLM proposal-and-review loop iteratively mutates known exploits (encoding, request placement) using only HTTP responses, with no visibility into WAF rules or Attack Score. Against an authorized customer staging environment, they ran 45 scenarios spanning XSS, SQL injection, command injection, SSRF, path traversal/LFI, and Log4j, recording 1,107 attempts. The vast majority were blocked; surviving requests were reviewed and turned into new detections. The exercise is becoming a foundational part of Cloudflare's WAF development lifecycle.
- LLM loop mutates payloads via real-time responses; models had no access to WAF rules
- 45 scenarios and 1,107 attempts across six attack categories including Log4j and SSRF
- Most attacks blocked; non-blocked requests led to new detections for all customers
- Tested WAF used Managed Ruleset, OWASP CRS Paranoia Level 3, Attack Score 30 blocking
From the original article
10 min read at blog.cloudflare.com
We built a WAF tester that adapted each request based on what the WAF blocked or passed. This helped us explore variations that a fixed test might miss. We ran it across six attack categories on an authorized staging environment and discovered detection gaps worth fixing.
By Vikram Grover · blog.cloudflare.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.