Cloudflare Connects OpenAI to Its Network for Faster Fixes

Cloudflare combines OpenAI models with its real-time traffic data to move from vulnerability detection to concrete protection proposals and fixes. Internally, the company is generalizing Codex among its developers, resulting in a reconstruction delivered in three days for $1,100, according to its management.

Generalized Codex and a Project Rebuilt in 3 Days for $1,100

Mark Anderson notes that one-third of the company consists of developers, and 100% of them use Codex to increase their output and improve code quality. He cites a reconstruction project carried out by a team of four developers, built in three days, deployed and tested, at a cost of $1,100 in tokens. Anderson estimates that, without this support, the project would have taken months of sequencing and perhaps ten times more developers. He links the use of Codex to increased productivity, better quality, and fewer bugs introduced.

A Customer Journey from Detected Vulnerability to Applied Protection

Cloudflare presents a structured approach that goes from detecting a vulnerability to securing the affected application. The service provides evidence to prioritize vulnerabilities, offers protective measures to limit exposure, and suggests code modifications for analysis, while maintaining control of changes with clients. This method is integrated into the early access service, which combines code analysis with evidence from the production environment and actions that need to be validated by clients. The stated goal is to provide security teams with the necessary information to prioritize, protect, and remediate, without taking away clients' control over deployments.

Code Analysis with Daybreak and Production Signals

The Vulnerability Discovery and Remediation service, within Cloudflare Managed Defense, combines OpenAI models accessible via Daybreak with the context of Cloudflare's global network. The models analyze the code that the client has authorized and identify potential vulnerabilities, while information from real-time traffic, security indicators, and validation operations help establish priorities among the results. The service develops suggestions for code fixes and, if necessary, custom WAF rules to limit exposure during developer reviews. These proposals undergo automated checks and are then reviewed by the Cloudflare team before being sent to clients.

Linking Code to Production to Sort and Act Faster

Cloudflare's approach involves linking each vulnerability in the source code to signals observed in production: checking if the affected route is active, if it is capturing suspicious traffic, and if existing controls already cover it. These signals add the necessary context to decide on priorities. Mark Anderson describes an environment where security teams receive a continuous stream of alerts, the manual processing of which is time-consuming and limits the ability to resolve critical cases due to a lack of resources. He argues that automation through Daybreak and Codex delegates basic tasks to machines, allowing human expertise to focus on key decisions. Meanwhile, Codex is used daily by engineering. Operationally, Cloudflare reminds that it proxies about 20% of web traffic, with clients relying on this network for security, performance, and reliability. For Anderson, integrating AI means reorganizing internal processes by considering machines as an integral part of the building cycle.