Cloudflare launches OpenAI-powered vulnerability defense tied to live traffic

Why it matters

Cloudflare is turning its edge telemetry into the context AI code scanners lack. If the checks hold up, its network becomes both the place that finds exposure and the fastest place to contain it.

Cloudflare, co-founded by Matthew Prince (@eastdakota), launched invitation-only early access to Vulnerability Discovery and Remediation on September 3rd that inspects customer-authorized code, connects vulnerabilities to production traffic and proposes both software patches and temporary firewall rules.

The launch pulls the company built by Matthew Prince, Michelle Zatlyn (@zatlyn) and Lee Holloway back to its founding idea: a security provider can make better decisions when it sees malicious behavior across many websites. The difference in 2026 is that OpenAI models are searching the code while Cloudflare supplies the network context.

Cloudflare calls the service Vulnerability Discovery and Remediation. It sits inside Cloudflare Managed Defense and uses OpenAI's GPT-5.6 Cyber for reconnaissance, vulnerability hunting and validation. Cloudflare then checks whether the affected code is deployed, which routes can reach it, how much traffic those routes carry, whether they have attracted suspicious requests and which Web Application Firewall protections already apply.

That combination is Cloudflare's central wager. AI models can produce vulnerability findings quickly. Security teams still have to decide which findings describe reachable, actively exposed code and which can wait. Cloudflare is selling the evidence required to make that decision.

From spam signals to runtime context

Prince and Holloway began building Project Honey Pot in 2004 to track how spammers harvested email addresses. Users eventually asked for a way to stop attackers rather than merely observe them. Prince brought the project to Harvard Business School, where Zatlyn recognized the commercial opportunity. Holloway built Cloudflare's first prototype while Prince and Zatlyn developed the business plan.

Vulnerability Discovery and Remediation follows the same progression from observation to intervention. Cloudflare sees requests passing through its network, WAF detections and the routes served by its Workers platform. The new service uses those records to decide where model-generated findings deserve attention, then prepares an edge control that can buy engineers time while they review a permanent code change.

Cloudflare's September 3rd press release describes the partnership and the service's edge-defense approach.

Cloudflare has been applying the same network-data advantage across a growing set of products. On August 31st, RuntimeWire reported that Cloudflare's bot detection engine continuously retrains on live traffic. The vulnerability service extends that strategy into application code: use activity observed at the edge to improve a decision that isolated software tools would otherwise make with incomplete information.

How the service works

An investigation begins with a snapshot from Cloudflare Web Assets and WAF. For a Cloudflare Worker, the service retrieves the latest authorized source version, maps its configured routes and matches those routes with request metadata from Workers Observability.

Cloudflare's reconnaissance agent maps request paths to the relevant sections of the codebase. Hunter agents inspect those sections, and a validation stage attempts to corroborate each finding in the source. Production evidence can raise a finding's risk rating when an affected endpoint carries substantial traffic or shows signs of active probing.

The output is a prioritized list with a recommended code patch. When Cloudflare says the evidence supports an edge mitigation, the service also prepares a narrowly scoped custom WAF rule. Customers that have authorized the service to defend a zone can have those rules deployed around the request method, path and other details needed to reach the vulnerable code.

Cloudflare says it will avoid proposing a rule when a route is too broadly defined to scope safely. Its engineers validate rule syntax and test suggestions against synthetic fixtures rather than customer traffic. Every result receives a human review, and customers decide whether to test or deploy the proposed changes.

The workflow grew out of an internal vulnerability harness Cloudflare described in June. That system separates reconnaissance, hunting and adversarial validation into persistent stages instead of asking one coding agent to hold an entire investigation in its context window. Cloudflare said repeated runs of its original single-session process found materially different sets of bugs, prompting it to build orchestration around the models.

The model runs at OpenAI

Despite the emphasis on Cloudflare's edge, inference does not run there. Cloudflare sends prompts through AI Gateway to OpenAI's servers. GPT-5.6 Cyber handles reconnaissance, hunting and validation, then returns its responses to Cloudflare's harness.

Cloudflare says each investigation is limited to the source and evidence authorized by the customer. It removes unnecessary context, applies engagement-specific redaction controls and logs tool calls. Mechanical checks outside the model can stop a patch or rule before it reaches the customer.

That architecture makes code access and model governance part of the purchasing decision. Customers are authorizing a managed investigation in which selected source code, logs and request metadata can be processed by an outside model provider. The early-access format gives Cloudflare's Managed Defense staff a role in setting those boundaries before the service reaches a broader market.

OpenAI introduced GPT-5.6 Cyber through its restricted Daybreak Red program on August 10th. OpenAI says the model is available for authorized vulnerability research, exploit validation and security testing.

GPT-5.6 Cyber is one component of Cloudflare's system, alongside route mapping, network evidence, mechanical checks and human review.

Cloudflare still has to prove the fixes

Cloudflare's announcement describes the operating controls in detail, but the commercial test will depend on the quality of the findings and patches. The hypothetical 4,000 vulnerabilities and 78 critical findings in its blog post are an illustration, rather than a customer result. The launch materials provide no independent precision, recall or remediation-success benchmark.

That evidence matters because AI vulnerability tools have struggled when moved from demonstrations into working development environments. A Microsoft Research user study found high rates of false positives and fixes that did not apply to the user's codebase. Researchers identified missing context and limited customization as central problems.

Cloudflare's design directly targets those weaknesses. It grounds model output in the customer's source, checks whether vulnerable routes are active and places model-generated changes behind external validation and human approval. A production case study will show whether those controls reduce noise without burying less obvious vulnerabilities.

Cloudflare already has a substantial enterprise channel for the experiment. Its 2025 annual filing reported 332,466 paying customers as of December 31st, including 4,298 customers generating more than $100,000 in annualized revenue. Early access is limited to selected enterprise customers, and each engagement starts with one authorized application.

The service also gives Cloudflare a way to bind products that customers may currently buy separately: Workers, Web Assets, observability, WAF and Managed Defense. That makes the launch a distribution play as much as a model integration. OpenAI supplies the specialized reasoning. Cloudflare supplies the production evidence, the mitigation point and the enterprise relationship.