Cloudflare ranks AI-found vulnerabilities by live production exposure

Cloudflare's answer to its own opening question (a scanner has flagged 4,000 vulnerabilities, 78 of them critical, so which comes first?) is to judge each finding by what its network sees in production. In the announcement, Cloudflare points out what a scanner cannot tell you: whether the flagged code is deployed, whether anyone is calling that route, what attack activity surrounds it, and which protections already cover it. Vulnerability Discovery and Remediation adds that context. The service is invitation-only, sold through Cloudflare Managed Defense, and works on Cloudflare Workers and proxied applications.

The service starts with a snapshot from Cloudflare's Web Assets and WAF (web application firewall) products, showing which routes are active, how much traffic they carry and whether security events are tied to them. Routes with heavy traffic are treated as hot paths, and the code behind them gets stricter scrutiny. An AI harness then works through the code the customer has authorized. A reconnaissance agent maps request paths to the code that handles them, and hunter agents search those sections for vulnerabilities. The models are OpenAI's Daybreak models, including GPT-5.6 Cyber, reached through OpenAI's Daybreak Defense Network. Cloudflare says network context can steer the hunters toward an active or recently targeted route, but it does not count as proof. Every finding has to be backed by evidence in the source code. A validation stage gives each finding an initial risk rating from the code. Traffic or signs of active probing on the affected endpoint can then raise it.

The result is a ranked list. Each finding comes with a proposed code patch and, where the evidence supports one, a narrowly scoped WAF custom rule that blocks requests able to reach the vulnerable code while the patch is reviewed. Cloudflare will not suggest a rule for a route pattern made up only of variables and wildcards. According to the post, the model can propose patches and rules but cannot apply them. Checks outside the model, plus validation by Cloudflare's team, must pass before anything reaches the customer. Customers decide what gets implemented, with one exception: a customer who has authorized the service to defend a zone can have Cloudflare deploy those scoped rules for them. Inference runs on OpenAI's servers, reached through Cloudflare AI Gateway, not on Cloudflare's edge. Cloudflare says it strips each investigation down to the authorized code and evidence and applies the redaction controls configured for the engagement before anything reaches the model.