Modern application security teams frequently face an overwhelming volume of alerts. When an automated vulnerability scanner flags thousands of new issues—including dozens categorized as critical—security professionals and software developers are immediately confronted with a daunting decision: which problem needs to be addressed first?
To help organizations navigate this growing challenge, Cloudflare has announced the early access launch of Vulnerability Discovery and Remediation, a new service integrated into Cloudflare Managed Defense. Designed to help customers detect and mitigate application-layer vulnerabilities within their authorized codebases, the service relies on an invitation-only model that brings advanced automation and deep network context to the prioritization and patching workflow.
Through participation in the OpenAI Daybreak Defense Network, Cloudflare leverages advanced AI models, including GPT-5.6 Cyber, to perform reconnaissance, threat hunting, and validation across codebases that customers explicitly authorize the company to access. When the system detects a potential vulnerability, it formulates proposed solutions. It then automatically checks each proposed patch and any accompanying mitigation strategies before presenting them to the customer for thorough review. Throughout this entire workflow, human operators and client teams remain firmly in control, deciding whether and when any code patches or mitigations are implemented.
The Growing Challenge of AI-Driven Discovery and Alert Fatigue
The difficulty of choosing what to fix first has historically been a persistent hurdle in application security, but the rapid evolution of artificial intelligence has significantly exacerbated the problem. Large language models are now capable of surfacing potential weaknesses across an entire enterprise codebase in mere minutes, causing the total number of findings to climb rapidly.
However, industry experts note that the core issue is no longer just the sheer volume of alerts, but the unprecedented speed of discovery. Malicious actors can similarly leverage automated AI tools to accelerate vulnerability reconnaissance and exploitation. This dynamic leaves security teams and internal developers with considerably less time to evaluate which findings actually pose a genuine risk and to act accordingly before an exploit occurs.
Traditional security scanners typically operate in an operational vacuum. For example, if a conventional scanner flags a vulnerability inside a specific code handler, it usually cannot determine whether that vulnerable code is actually deployed to a production environment. It cannot verify whether real-world traffic is hitting that specific route, what security telemetry surrounds it, or what defensive controls are already active in the network infrastructure. Consequently, security teams are frequently forced to prioritize raw findings without concrete evidence of production exposure or existing compensatory protections.
Combining Global Network Context with Source Code Analysis
Cloudflare aims to bridge this visibility gap by leveraging its expansive global network. Because the company sits in the data path for a massive share of internet traffic, its infrastructure can observe which application routes are actively receiving requests, how much traffic those routes carry, and what kinds of security events or suspicious activities surround them.
When customers enable Vulnerability Discovery and Remediation alongside Cloudflare’s Web Application Firewall (WAF), the service gains visibility into which security rules are already deployed and actively blocking attacks in real time. This contextual data transforms a generic, abstract vulnerability finding into a highly specific priority. Instead of an unverified alert, security analysts can see that a vulnerability exists in live code on a heavily utilized route, complete with recent probing activity and no existing compensating protection.
To address these high-risk areas, the service can help mitigate the exposure by proposing custom WAF rules and tailored code patches designed specifically for the customer’s systems. This approach mirrors the internal methodologies Cloudflare previously described in its technical disclosures regarding its model-agnostic vulnerability scanning pipeline, which scans the company’s own infrastructure fleet, adversarially validates every finding, and translates raw AI model outputs into actionable fixes for internal engineers.
How the Vulnerability Harness Works in Production
Operating across Cloudflare Workers and proxied web applications, the vulnerability discovery process begins by gathering a comprehensive snapshot of traffic and security data from Web Assets and the WAF. This snapshot establishes a baseline of active routes, traffic volumes, and associated security triggers. Pathways exhibiting high volumes of detection triggers receive heightened scrutiny during the security profiling process.
For serverless applications built on Cloudflare Workers, the system retrieves the most recent source version and its configured routes to map out the exact endpoints served. By matching these routes against Web Assets and request metadata derived from Workers Observability, the service ties the specific source code under review directly to the endpoints it handles in production. This collected network context remains accessible to the underlying AI agents throughout the investigation.
The vulnerability harness then initiates its investigative workflow. A specialized reconnaissance agent maps request paths to the corresponding sections of the authorized codebase. Armed with this map, hunter agents investigate specific parts of the code to uncover potential vulnerabilities, pulling in relevant network context as needed. While this context helps the hunter agents focus their attention on code behind active or recently targeted routes, it does not automatically establish that a vulnerability exists; every finding must be corroborated by concrete evidence within the source code itself.
Once the hunter agents complete their analysis, a validation stage evaluates the proposed mitigations and assigns an initial risk rating based strictly on the source code. The collected network evidence can subsequently elevate that risk rating further if the affected endpoint carries significant production traffic or demonstrates active probing attempts.
The culmination of this process is a prioritized list of findings, accompanied by recommended code patches and, where supported by sufficient evidence, a custom Cloudflare WAF rule designed to reduce exposure while the code fix undergoes human review. If the customer has authorized the service to defend their zone, these rules are deployed with conservative scoping around the specific HTTP methods, paths, and request details required to reach the vulnerable code. To minimize false positives, if a route pattern contains only variables and wildcards without explicit path identifiers, the system deliberately refrains from suggesting an edge rule, prioritizing accuracy over speculative protection.
Model Execution and Safety Boundaries
When an authorized investigation is initiated, the vulnerability harness runs within Cloudflare’s infrastructure, while model prompts are transmitted from Workers through Cloudflare AI Gateway to OpenAI Daybreak models hosted on OpenAI servers. Specifically, GPT-5.6 Cyber is utilized during the reconnaissance, hunting, and validation phases, with its responses routed back to the harness to continue the workflow on Cloudflare. Importantly, no model inference occurs directly at Cloudflare’s edge, and the AI models themselves lack the capability to independently apply any proposed patches or firewall rules.
To maintain strict operational security, each investigation is tightly scoped to the source code and evidence authorized by the customer. Prior to reaching the model, the system strips out unnecessary data and applies configured redaction controls. The underlying harness treats source code, application logs, and request metadata strictly as evidence to be inspected rather than instructions to be executed.
Tool access is similarly governed by strict operational boundaries. Every tool call is logged and verified against the engagement’s access policy before execution, and every proposed patch or rule must pass automated checks implemented completely outside the model environment. If any validation check fails, the workflow halts immediately before a proposal can reach customer review.
Nothing is presented to the client team until it has successfully cleared all automated safety checks and been validated by Cloudflare’s security personnel. For edge-defense suggestions, this involves validating rule syntax and running the rules against synthetic fixtures representing expected traffic patterns, rather than testing them directly against live customer traffic. If a check fails or the result remains ambiguous, the system withholds the output and routes it for manual diagnostic review. Even after successful validation, the service merely prepares the source code patch and WAF rule, leaving all final implementation decisions entirely in the hands of the customer.
Early Access Availability
Vulnerability Discovery and Remediation is currently being offered to select customers through an invitation-only early access period managed by Cloudflare’s Managed Defense team. Each individual engagement begins with a single application whose codebase the customer formally authorizes Cloudflare to investigate.
To successfully correlate findings with live production traffic, the service utilizes authorized read access to the client’s Web Assets operation inventory, relevant WAF controls, and Workers Trace Events Logpush data where available. While the investigation workflow is semi-automated, human oversight remains central, ensuring that customer engineering teams review every single result before deciding whether to test or deploy any suggested changes. Customers interested in participating in the early access program are advised to reach out directly to their Cloudflare account teams for further information.
Leave a Reply