Cloudflare Tests WAF Against AI-Generated Attacks, Enhancing Defenses
Cloudflare has rigorously tested its Web Application Firewall (WAF) against sophisticated, AI-generated attack payloads, demonstrating its effectiveness while identifying areas for continuous improvement.
Cloudflare is proactively addressing the evolving threat landscape by testing its own Web Application Firewall (WAF) against advanced AI-generated attack payloads. In response to customer inquiries about the readiness of WAFs against frontier AI models, Cloudflare developed a novel testing methodology. This approach simulates a hacker's perspective, allowing an AI model to dynamically iterate and mutate exploit attempts without prior knowledge of the WAF's internal rules or the application's source code.
The core of this testing framework involves a Python-based WAF tester that begins with known exploits. It then employs Large Language Models (LLMs) to systematically alter these exploits by changing encoding, delivery methods, or payload variations. The system analyzes the HTTP responses to these mutated attempts, using the feedback to guide the next iteration. This adaptive loop allows the AI to explore a vast attack surface much faster than human testers could, mimicking the agility of sophisticated adversaries.
Cloudflare's dynamic testing approach contrasts with traditional static and dynamic application security testing. Instead of analyzing code or probing running applications with predefined scripts, the LLM acts as an intelligent agent. It receives limited information—primarily selected HTTP response data—and uses this to refine its attack strategies. This method is designed to uncover vulnerabilities that might be missed by less adaptive security tools, providing a more realistic assessment of WAF efficacy.
During the main test, Cloudflare ran 45 scenarios against an authorized customer staging environment protected by its WAF. These scenarios covered six critical attack categories: cross-site scripting (XSS), SQL injection (SQLi), command injection (CMDi), server-side request forgery (SSRF), path traversal/local file inclusion (LFI), and Log4j exploits. The WAF was configured with specific blocking thresholds and enabled managed and OWASP Core Rulesets at a high paranoia level.
Across 1,107 recorded attempts, the Cloudflare WAF successfully blocked the vast majority of AI-generated attacks. However, the requests that managed to bypass the WAF were invaluable. These successful, albeit limited, bypasses were meticulously reviewed by human analysts. After filtering out malformed, benign, duplicate, and out-of-scope observations, the remaining unblocked requests provided crucial insights for enhancing detection capabilities.
The findings from these bypasses directly informed the creation of new detection rules, further hardening Cloudflare's security posture for all its customers. This continuous feedback loop, where AI-driven testing identifies weaknesses that are then patched, is becoming a foundational element of Cloudflare's WAF development lifecycle. It ensures that defenses evolve in lockstep with the increasingly sophisticated threats posed by AI.
Cloudflare emphasizes that even a WAF bypass is not a complete compromise. A successful attack still requires an underlying exploitable vulnerability in the application itself. Therefore, the company strongly advises organizations to maintain a robust patch management strategy for their software. Keeping applications up-to-date remains one of the most effective defenses against attackers, complementing the layered security provided by advanced WAF solutions.
This initiative underscores Cloudflare's commitment to staying ahead of emerging threats. By leveraging frontier AI models to test and improve its own security products, the company aims to provide robust protection against the next generation of cyberattacks, ensuring that its customers' applications remain secure in an increasingly complex digital environment.