Cloudflare Correlates Code Vulnerabilities With Live Traffic to Generate Patches and WAF Rules Through a Multi-Agent Workflow

Cloudflare has launched Vulnerability Discovery and Remediation (VDR) as an invitation-only offering within Managed Defense, aiming to address traditional SAST tools’ and vulnerability scanners’ lack of production context. The service first collects routes, traffic, recent security events, and existing protections from Web Assets, WAF, and Workers Observability, then correlates these signals with source code that the customer has authorized it to inspect. Findings involving high-traffic endpoints or endpoints currently being probed can receive higher priority, but network activity alone is not treated as evidence of a vulnerability.

The analysis is orchestrated by a multi-stage harness: a reconnaissance agent first maps HTTP paths to handlers, hunter agents search separate areas for weaknesses, and a validation stage then confirms issues using code evidence, checks proposed fixes, and assigns scores. The system can propose both a long-term code patch and a WAF Custom Rule that reduces exposure until the fix is deployed. If a route can only be described using a wildcard and the attack surface cannot be conservatively constrained, the system does not recommend generating a rule.

The design’s trust boundary warrants attention. The harness runs on Cloudflare, but Workers send the necessary prompts, code, and evidence through AI Gateway to the Daybreak model on OpenAI’s servers. GPT-5.6 Cyber is used for reconnaissance, hunting, and validation; inference does not run at the Cloudflare edge. Cloudflare says it removes irrelevant content, applies agreed-upon redaction rules, and treats code and logs as data rather than instructions. Every tool call is also constrained by policy checks and logging mechanisms outside the model.

Under the standard workflow, findings are delivered to the customer for review after validation by Cloudflare personnel; the model does not apply patches or rules on its own. Customers can separately authorize the service to deploy conservatively scoped WAF protections. Engineering teams should next focus on detection recall, false positives, test coverage for generated fixes, and data-retention and audit requirements spanning Cloudflare and OpenAI. The service is currently available only to select enterprise customers, beginning with one authorized application at a time, and Cloudflare has not provided independently reproducible performance benchmarks.