Cloudflare Launches AI-Powered Vulnerability Defense Service with OpenAI

Cloudflare has launched a new service that integrates OpenAI's cybersecurity model into its security monitoring platform, enabling the identification of code vulnerabilities and prevention of attacks at the edge. The service aims to accelerate vulnerability detection through AI, prioritizing items with higher risk in real operational environments.
The company announced on September 3 that it is offering the 'Vulnerability Discovery and Response' service as part of its Managed Defense suite, available through early access for select Cloudflare Enterprise customers in San Francisco. Cloudflare disclosed in its announcement that as of September 2024, there were 64,750 vulnerabilities registered in the U.S. National Vulnerability Database (NVD), surpassing the total of 48,185 registrations projected for 2025.
Cloudflare explained that traditional vulnerability scanners often present thousands of findings without operational context, causing security teams to spend excessive time determining actual risks. The new service uses OpenAI's Breakout Defense Network with GPT-5.6 Cyber models to analyze codebases authorized by customers. It connects the analysis to real traffic patterns, security events, and edge control features.
Unlike conventional scanners that merely list vulnerabilities, Cloudflare first creates snapshots of web asset inventories and web application firewall (WAF) traffic and security data. It identifies which paths are actively open and whether security events have occurred on those paths before narrowing the code investigation scope. The reconnaissance agent links request paths to specific code segments, while the exploration agent searches for vulnerabilities within authorized code sections, assigns risk ratings after verification.
Vulnerabilities with high traffic volume or confirmed exploration attempts in operational environments are prioritized. In the Workers environment, it matches recent source versions and configured routes to actual endpoints, using Workers Observability request metadata to confirm whether code weaknesses align with live service paths.
Cloudflare clarified that model inference does not occur at the edge but is executed within Cloudflare's workflows after customer approval. Model prompts are sent via an AI gateway through Workers to OpenAI servers, with responses returning to Cloudflare's workflow. All proposed patches and firewall rules require human review before implementation, as the scope of investigation remains limited to authorized code and evidence.
Matthew Prince, co-founder and CEO of Cloudflare, stated that security teams are losing if they manually counter AI-driven attacks: 'We need to shift from patching vulnerabilities one by one to an automated defense system.' McCall McIntyre, OpenAI's Global Cyber Partnership Lead, emphasized that the Breakout Defense Network aims to safely provide frontier AI benefits to defenders while limiting high-performance models to approved environments for defensive purposes.
OpenAI previously disclosed in its August 10 Breakout documentation that GPT-5.6 Cyber is available through the Breakout Red access tier, trained specifically for advanced cybersecurity tasks like zero-day vulnerability discovery and exploit chain development. The company reported a 95.0% completion rate in internal high-risk security evaluations for GPT-5.6 Cyber, compared to 1.5% for GPT-5.6 Sol, 2.0% for GPT-5.6 Sol with Breakout Blue access, and 57.3% for GPT-5.5 Cyber. OpenAI noted that GPT-5.6 Cyber is rated 'high' in cybersecurity capability within its framework but has not yet reached the 'critical' threshold.
The service focuses on prioritizing vulnerabilities based on real traffic and security events rather than merely detecting them, which aligns with concerns in crypto sectors like exchanges, wallets, payment services, and open-source libraries where code weaknesses could lead to asset losses. Cloudflare's approach of requiring human approval for patches and firewall rules addresses ongoing debates about restricting access to high-performance AI models for defensive use.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: 클라우드플레어, 오픈AI로 취약점 방어 서비스 시작