Cloudflare is opening early access to Vulnerability Discovery and Remediation, a new invitation-only service under Cloudflare Managed Defense designed to help customers detect and fix vulnerabilities in their codebases. The service uses OpenAI’s Daybreak models, including GPT-5.6 Cyber, running through what Cloudflare calls the OpenAI Daybreak Defense Network, to perform reconnaissance, hunting, and validation against codebases customers explicitly authorize.
When a vulnerability turns up, Cloudflare proposes solutions and automatically checks each patch or mitigation before presenting it for review, though customers retain final say over whether anything actually gets implemented.
The service addresses a growing prioritization problem. AI tools have made it faster than ever to surface potential weaknesses across a codebase, which means the sheer volume of findings keeps climbing, but speed cuts both ways since attackers can also use AI to accelerate vulnerability discovery and exploitation, leaving security teams less time to figure out what actually matters. A typical scanner might flag a vulnerability in a handler without indicating whether that code is even deployed, whether anyone’s actually hitting that route, or what protections are already in place, leaving teams to prioritize blind.
Cloudflare’s pitch is that its network visibility fills in that gap. By seeing which routes are active, how much traffic they carry, and what security events surround them, and additionally seeing which WAF rules are already blocking attacks when Web Application Firewall is enabled alongside the service, Cloudflare can turn a generic finding into a specific, evidence-backed priority: a vulnerability that’s live in production, on a heavily used route, with recent attack activity and no existing protection. From there, it can propose custom WAF mitigations and code patches tailored to the customer’s systems.
The underlying approach builds on a model-agnostic vulnerability-hunting pipeline Cloudflare previously used to scan its own infrastructure, adversarially validating findings and turning raw model output into fixes engineers can trust. Vulnerability Discovery and Remediation extends that internal system to customer-authorized code, then connects the findings to production traffic, security events, and the edge controls that can act on them.
The detection process starts by pulling a snapshot of traffic and security data from Web Assets and WAF, showing which routes are active, how much traffic they see, and whether recent security events are tied to them; routes carrying heavy traffic get flagged as hot paths and undergo stricter security profiling. Source code vulnerability analysis then identifies potential weaknesses, but on its own doesn’t reveal which routes reach that code or how exposed it actually is, so Cloudflare combines both signals. For Workers specifically, the service pulls the most recent source version along with configured routes, matching them against Web Assets and request metadata from Workers Observability to tie exact source code to the production endpoints it serves.
From there, a Reconnaissance agent maps request paths to the relevant parts of the codebase, then sends hunter agents into those sections to look for vulnerabilities, pulling in network context as needed, though every finding still needs to be corroborated by evidence in the actual source code. A validation stage then checks proposed mitigations and assigns each vulnerability an initial risk rating based on the code itself, which can be raised further if network evidence shows the affected endpoint carries heavy traffic or shows signs of active probing.
The end result is a prioritized list of findings, each paired with a recommended code patch and, where the evidence supports it, a Cloudflare WAF custom rule that can reduce exposure while the underlying code fix goes through review. If a customer has authorized the service to defend their zone, Cloudflare deploys these rules scoped conservatively around the specific method, path, and request details needed to reach the vulnerable code, and deliberately avoids suggesting a rule if a route pattern is too broad or made up mostly of variables and wildcards.
On where the actual AI processing happens, Cloudflare runs the investigation harness on its own infrastructure but sends model prompts through Cloudflare AI Gateway to OpenAI’s Daybreak models running on OpenAI’s servers, with GPT-5.6 Cyber used for reconnaissance, hunting, and validation before responses return to Cloudflare’s systems. No model inference happens at Cloudflare’s edge, and the model itself cannot apply any patch or rule it proposes. Each investigation is scoped narrowly to only the source code and evidence a customer authorizes, with unnecessary context redacted before it ever reaches the model, and source code, logs, and request metadata are treated strictly as evidence for the model to inspect rather than instructions to follow.
Every tool call is logged and checked against the investigation’s access policy, and every proposed patch or rule must pass checks implemented outside the model itself before it can reach customer review; for WAF rule suggestions specifically, that includes validating rule syntax against synthetic test traffic rather than real customer data. If any check fails or a result stays ambiguous, the output is held back for further diagnosis rather than surfaced to the customer, and passing those checks still doesn’t change anything in the customer’s environment automatically, since the final patch and rule are only prepared for review after Cloudflare’s own team validates the output.
