Cloudflare tested its WAF with frontier models and reviewed the apparent bypasses

The exercise used 1,107 attempts against an authorized staging environment. Requests that passed the firewall became investigation leads, not automatic exploit findings.

Por Spargax team

29 de septiembre de 20261 min de lectura

Actualizado 18 de octubre de 2026

Cloudflare used frontier models to vary known attack inputs and test its web application firewall against an authorized customer staging environment. Its September 29 report covers 1,107 attempts across six attack categories.

The model could see selected HTTP responses but not the application’s source code or the firewall’s rules. It changed how an input was encoded or delivered, then used the response to choose another variation.

A request getting through the WAF was only a lead. Human reviewers removed malformed, benign, duplicate and out-of-scope observations before deciding which results exposed a detection gap. Cloudflare says the remaining findings informed new detections.

The exercise separates a firewall bypass from a successful attack on the application behind it. An accepted request still needs an exploitable application behavior to cause the intended harm. The reported results remain specific to the authorized environment and inputs used.

Fuentes