Summary: Cloudflare announced early access to Vulnerability Discovery and Remediation on September 3, 2026 as part of Managed Defense. The service uses OpenAI Daybreak models, including GPT-5.6 Cyber, to hunt findings and connect them to production routes, traffic, security events, and existing WAF protections. It can propose patches and WAF rules, but model checks and human review occur before any customer deployment.
A scanner reports 4,000 fresh vulnerabilities, 78 of them critical. Where do you start? That question was never easy, and it keeps getting harder. Large language models can comb a codebase and surface weaknesses in minutes, so findings keep piling up. Attackers, meanwhile, lean on AI to speed up their own discovery and exploitation, which leaves security teams less time to work out what actually matters.
Cloudflare announced early access to Vulnerability Discovery and Remediation (VDR) on September 3, 2026, as part of Cloudflare Managed Defense. Built on OpenAI Daybreak models, GPT-5.6 Cyber among them, the service hunts for vulnerabilities in the codebases you authorize. What sets it apart is context: every finding gets tied to production traffic, security events and existing protections, so you can tell which vulnerability is genuinely live and exposed.
Why context matters
A conventional scanner will tell you a handler carries a vulnerability. It won’t tell you whether that code is deployed, whether any traffic reaches the route, or which security controls already sit around it. Prioritization then rests on feel rather than evidence of production exposure.
Cloudflare’s global network closes that gap. Turn on VDR alongside Web Application Firewall (WAF), and Cloudflare can observe which routes are active, how much traffic each one carries and which security events attach to them. It also sees which WAF rules are already applied and actively blocking attacks. A generic finding turns into a concrete priority: live code, a heavily used route, recent attack activity, no protection in place.
None of this is entirely new. In an earlier post, Cloudflare laid out the model-agnostic pipeline it runs against its own fleet — scanning for issues, adversarially validating findings, and turning raw model output into fixes engineers can trust. VDR carries that internal harness over to customer code, then links its findings to production signals and edge controls.
How the harness works
Everything opens with a snapshot of traffic and security data pulled from Web Assets and WAF. It shows which routes are active and whether security events have recently touched them. Routes carrying high request volume count as “hot paths” and receive stricter security profiling.
On Cloudflare Workers, VDR fetches the newest source version together with its configured routes, then lines those routes up against request metadata from Workers Observability. The exact source under review is thereby tied to the endpoints it serves in production. That network context stays reachable for the whole investigation.
Then the vulnerability harness runs. A Reconnaissance agent maps request paths onto the code that handles them. Hunter agents move into specific sections, searching for vulnerabilities and drawing in network context where it helps. Context alone never proves a vulnerability exists — every finding has to be corroborated by evidence in the source code.
Once the hunters report back, a validation stage examines proposed mitigations and sets an initial risk rating from the source code. Network evidence can push that rating higher — a heavily trafficked endpoint or signs of active probing, for instance.
What comes out is a prioritized list. Each finding carries a recommended code patch and, where the evidence allows, a WAF Custom rule that lowers exposure while the code fix moves through review. Cloudflare proposes no rule when a route pattern consists only of variables and wildcards — missing a possible connection is preferable to asserting one the evidence cannot support.
Where the model runs and what it can’t do
Once you authorize an investigation, VDR runs the harness on Cloudflare and routes model prompts from Workers through Cloudflare AI Gateway to OpenAI Daybreak models running on OpenAI’s servers. Reconnaissance, hunting and validation all use GPT-5.6 Cyber. No model inference happens at Cloudflare’s edge, and no model can apply a patch or rule of its own accord.
Investigations stay narrow by design. Before any context reaches the model, VDR strips what the investigation doesn’t require and applies redaction controls. Source code, logs and request metadata enter the harness as evidence to inspect, never as instructions to follow. Tool access obeys the same boundary: every call is logged and checked against an access policy before it runs. Each patch or rule proposal has to clear checks implemented outside the model, and a failed check halts the workflow before the proposal ever reaches customer review.
Nothing reaches you until those checks pass and Cloudflare’s own team has validated the output. For a WAF rule, that means validating syntax and running it against synthetic fixtures rather than customer traffic. Clearing the checks still leaves your environment untouched — every result is yours to review before you decide whether to test or deploy.
Early access and practical next steps
During early access, VDR is invitation-only and arranged through Cloudflare’s Managed Defense team. An engagement opens with a single application you authorize. The investigation is semi-automated, yet control stays with you: Cloudflare may propose patches and mitigations, but implementing them is your decision.
If you’re a Cloudflare customer watching a growing pile of scanner findings, this service may help you sort the critical and exposed from the noise. To learn about joining early access, talk to your Cloudflare account team.