Your scanner just flagged 4,000 new vulnerabilities, 78 of them critical. Which one do you fix first? That question has always been hard, but it’s getting harder. Large language models can surface weaknesses across a codebase in minutes, so the number of findings keeps climbing. Meanwhile, attackers are using AI to accelerate their own discovery and exploitation, leaving security teams less time to decide what matters.
On September 3, 2026, Cloudflare announced early access to Vulnerability Discovery and Remediation (VDR), part of Cloudflare Managed Defense. The service uses OpenAI Daybreak models, including GPT-5.6 Cyber, to hunt for vulnerabilities in codebases you authorize. But what makes it different is context: it ties each finding to production traffic, security events, and existing protections, so you know which vulnerability is actually live and exposed.
Why context matters
A typical scanner tells you there’s a vulnerability in a handler, but not whether that code is deployed, whether anyone is hitting that route, or what security controls already surround it. You’re left to prioritize without evidence of production exposure.
Cloudflare’s global network changes that. When you enable VDR with Web Application Firewall (WAF), Cloudflare can see which routes are active, how much traffic they carry, and what security events are associated with them. It can also see which WAF rules are already applied and actively blocking attacks. That context turns a generic finding into a specific priority: this vulnerability is in code that’s live, on a heavily used route, with recent attack activity and no existing protection.
This isn’t a brand-new idea. In a previous post, Cloudflare described the model-agnostic pipeline it uses to scan its own fleet, adversarially validate findings, and turn raw model output into fixes engineers can trust. VDR brings that internal harness to customer code, then connects findings to production signals and edge controls.
How the harness works
The process starts with a snapshot of traffic and security data from Web Assets and WAF. That shows which routes are active and whether they’ve seen recent security events. Routes with high request volume are treated as “hot paths” and get stricter security profiling.
For Cloudflare Workers, VDR retrieves the latest source version and its configured routes, then matches those routes to request metadata from Workers Observability. This ties the exact source under review to the endpoints it handles in production. The network context stays available throughout the investigation.
The vulnerability harness then runs. A Reconnaissance agent maps request paths to the code that handles them. Hunter agents then dig into specific sections, looking for vulnerabilities and pulling in relevant network context as needed. But context doesn’t prove a vulnerability exists—every finding must be corroborated by evidence in the source code.
After hunters return, a validation stage checks proposed mitigations and assigns an initial risk rating based on source code. Network evidence can raise that rating if, for example, the endpoint carries significant traffic or shows signs of active probing.
The result is a prioritized list of findings, each with a recommended code patch and, when evidence supports it, a WAF Custom rule that can reduce exposure while the code fix is reviewed. If a route pattern contains only variables and wildcards, Cloudflare won’t suggest a rule—it would rather miss a possible connection than claim one the evidence can’t support.
Where the model runs and what it can’t do
When you authorize an investigation, VDR runs the harness on Cloudflare and sends model prompts from Workers through Cloudflare AI Gateway to OpenAI Daybreak models on OpenAI’s servers. GPT-5.6 Cyber is used for reconnaissance, hunting, and validation. No model inference runs at Cloudflare’s edge, and the model cannot apply any patch or rule it proposes.
Each investigation is kept narrow. Before context reaches the model, VDR removes what the investigation doesn’t need and applies redaction controls. The harness treats source code, logs, and request metadata as evidence to inspect, not instructions to follow. Tool access is logged and checked against an access policy before each call runs. Every patch or rule proposal must pass checks implemented outside the model; if a check fails, the workflow stops before the proposal reaches customer review.
Nothing is presented until it clears those checks and Cloudflare’s team validates the output. For a WAF rule, that means validating syntax and running it against synthetic fixtures, not customer traffic. Passing checks still doesn’t change your environment—you review every result before deciding whether to test or deploy.
Early access and practical next steps
Vulnerability Discovery and Remediation is invitation-only during early access through Cloudflare’s Managed Defense team. Each engagement starts with one application you authorize. The investigation is semi-automated, but you stay in control: Cloudflare may propose patches and mitigations, but you decide whether they’re implemented.
If you’re a Cloudflare customer facing a growing pile of scanner findings, this service could help you separate the critical, exposed vulnerabilities from the noise. Talk to your Cloudflare account team to learn more about joining early access.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.