Cloudflare has begun offering early access to a service called “Vulnerability Discovery and Response” that finds vulnerabilities in customers’ application software and blocks attacks at the network edge.

According to a recent report by SiliconANGLE, the service uses OpenAI’s GPT-5.6-Cyber model via the Daybreak Defense Network and runs within Cloudflare’s security monitoring service, Managed Defense.

Cloudflare first identifies the paths that are actually open and the security events for those paths, based on its web asset inventory and web application firewall information. A reconnaissance agent then links request paths with segments of the codebase, and a detection agent uses that to find vulnerabilities. Code running on Cloudflare Workers is analyzed through Workers Observability.

Detection results are verified and assigned risk ratings. Risk increases when there is heavy traffic to the path in the production environment or when actual probing attempts are confirmed.

Firewall rules and code patches must all be approved by humans, and the model cannot apply them on its own.

OpenAI in August split its Daybreak access system, which allows some preselected organisations to access its advanced models, into two stages. It placed GPT-5.6-Cyber in Daybreak Red, a higher tier for vulnerability research and exploit verification.