Cloudflare connects vulnerability analysis with observable production risk

Cloudflare has announced early access to Vulnerability Discovery and Remediation, a Managed Defense service that addresses a practical problem with modern scanners: finding more vulnerabilities does not automatically reveal which ones should be fixed first. The workflow connects analysis of customer-authorized code with production signals such as active routes, traffic volume, security events, and existing WAF rules.

The distinction matters. Static analysis may identify a weakness in a handler, but it does not necessarily know whether that code is deployed, whether the route receives traffic, or whether an existing protection already blocks the relevant attempts. Cloudflare first collects a snapshot from Web Assets and the WAF. For Workers, it also connects the latest source version to configured routes and Workers Observability data. This creates a map between real requests and the parts of the code that handle them.

The process uses several agents with separate responsibilities. A reconnaissance agent identifies which routes reach which parts of the code, while other agents investigate specific areas. Traffic and suspicious-activity context helps prioritize the investigation, but it is not treated as proof that a vulnerability exists. Cloudflare says every finding must be corroborated by evidence in the source code. A validation phase then reviews proposed mitigations and assigns an initial priority based on the code; production signals can raise that priority when they show significant exposure or attack activity.

The system can propose a patch and, where appropriate, a custom WAF rule limited to the method, route, and request details needed to reach the vulnerable code. Rules are tested against synthetic fixtures representing expected requests before they are presented for review. Cloudflare also says that the model cannot apply the patch or rule by itself, and that the customer decides whether changes are implemented within the authorized workflow.

Architecturally, the model acts as a bounded analyst rather than a final authority. Access is limited to the code and evidence authorized for the investigation, redaction controls are applied, and every tool call is logged and checked against an access policy. Source code, logs, and request metadata are treated as evidence to inspect, not instructions to follow.

The lesson extends beyond this product. Security prioritization improves when it combines technical severity, real exposure, and existing controls, but those signals must remain separate from proof of the flaw. An agent can help investigate and prepare a defense; verifiable evidence, deterministic checks, least privilege, and human review remain the boundaries that turn an automated suggestion into a safe operational decision.