Cloudflare taps OpenAI GPT-5.6 Cyber to rank vulnerabilities by real production exposure

Cloudflare has announced early access to Vulnerability Discovery and Remediation (VDR), an invitation-only service inside Cloudflare Managed Defense that hunts for weaknesses in customer-authorized codebases and ranks them by real-world exposure. According to the Cloudflare blog, the service leans on OpenAI Daybreak models — including GPT-5.6 Cyber, reached through the OpenAI Daybreak Defense Network — for reconnaissance, hunting and validation.

The problem VDR targets is triage. A scanner can flag a flaw in a handler but cannot say whether that code is deployed, whether anyone hits the route, or which protections already sit in front of it. Cloudflare's pitch is that its network can answer those questions: through Web Assets, its WAF and Workers Observability, it can see which routes are live, how much traffic they carry and what security events surround them, including which WAF rules are actively blocking attacks.

How the pipeline works

VDR covers both Cloudflare Workers and proxied applications. An investigation starts by pulling a traffic and security snapshot that identifies active routes, request volumes and recent detection triggers. Routes with heavy traffic are treated as "hot paths", and the code deployed to them undergoes stricter security profiling. For Workers, the service fetches the latest source version and its configured routes, then ties that exact source to the endpoints it serves in production.

The vulnerability harness then takes over. A reconnaissance agent maps request paths to the sections of code that handle them, and hunter agents are dispatched into those sections of authorized code. Hunters can consult the collected network context — for example, to focus on code behind a route that is active or recently targeted — but context alone never establishes that a vulnerability exists; every finding must be corroborated by evidence in the source code. A validation stage then checks the proposed mitigations and assigns each vulnerability an initial risk rating based on the code, which network evidence such as heavy traffic or signs of active probing can raise further.

The output is a prioritized list of findings, each paired with a recommended code patch and, where the evidence supports it, a WAF Custom rule scoped narrowly around the HTTP method, path and other request details needed to reach the vulnerable code. Cloudflare says it declines to propose a rule when a route pattern consists only of variables and wildcards, preferring to miss a possible connection over claiming one the evidence cannot back. If a customer authorizes VDR to defend their zone, the rules are deployed to cut exposure while engineers review the patch — but the final decision on any patch or mitigation stays with the customer.

Cloudflare notes the service builds on work described in an earlier post, "Build your own vulnerability harness", which covered the model-agnostic pipeline it uses to scan its own fleet with adversarial validation. That internal system is one pillar of VDR; the new service extends it to customer code and connects the findings to production traffic, security events and the edge controls that can act on them.

Where the model runs

No model inference happens at Cloudflare's edge. Prompts are sent from Workers through Cloudflare AI Gateway to OpenAI Daybreak models running on OpenAI's servers, and the responses return to the harness on Cloudflare. The model cannot apply any patch or rule it proposes.

Each investigation is bounded to the source code and evidence the customer authorizes. Before context reaches the model, VDR removes what the investigation does not need and applies the redaction controls configured for the engagement. The harness treats source code, logs and request metadata as evidence to inspect rather than instructions to follow, which limits the reach of any directives embedded in the material being scanned. Every tool call is logged and checked against the investigation's access policy.

Why it matters

Two trends make triage the bottleneck in application security. Large language models can now surface weaknesses across a codebase in minutes, so the volume of findings keeps climbing; at the same time, attackers can use AI to accelerate discovery and exploitation, shrinking the window in which teams must decide what matters and act. Joining code-level findings with production traffic, security events and existing edge controls turns a generic finding into a specific priority — live code, on a heavily used route, with recent attack activity and no protection in place — and pre-validated patches with scoped WAF rules shorten the path from detection to mitigation. VDR is also an early example of a major infrastructure provider wiring third-party frontier cyber models into a commercial security product while keeping inference off its own edge and a human approval gate in front of every change, an architecture other vendors are likely to watch closely.