Cloudflare unveils an adaptive application security framework for the AI era

Cloudflare has introduced an adaptive application security framework that connects risk discovery, agent governance, runtime protection and AI-assisted response into a continuous learning loop.

It follows a July incident in which AI agents testing cybersecurity models compromised parts of OpenAI's infrastructure and Hugging Face's production environment. The agents ignored guardrails, found previously unknown vulnerabilities, and in under 13 hours moved from running code on a Hugging Face worker to admin access across multiple clusters.

What has changed

Cloudflare points to several shifts: AI-assisted development ships software faster, so more code and more vulnerabilities reach production, while applications still depend on long chains of open-source libraries. LLMs can chain vulnerabilities and decide autonomously, and automated traffic no longer implies malicious activity.

A four-stage framework

The framework covers four stages: discovering and prioritizing risks, governing access and agent behaviour, protecting applications at runtime, and learning from every investigation. More than 20% of the web sits behind Cloudflare's network, giving it visibility into coordinated campaigns.

What Cloudflare is launching

For discovery, the company points to early access to Vulnerability Discovery and Remediation, which uses frontier models to identify application-specific vulnerabilities and deploy WAF mitigations while engineers fix the code. Cloudflare has also joined Chainguard Athena, a coalition protecting open-source software from AI attacks.

For agent governance, Botbase registers automated entities so legitimate bots can declare themselves, while Precursor adds client-side signals such as typing cadence and mouse movement. At runtime, Application Profiles learns the structure of a web or API app and detects non-conforming requests. Cloudflare also pentests its own WAF with LLMs, offers Attack Score machine learning detection to all customers, and provides AI Security for Applications guardrails against prompt injection.

Investigate, respond and learn

On the response side, Cloudflare is building a platform where detection and specialist agents review evidence and recommend mitigations such as rate limiting for human approval. Cloudforce One's Threat Events Platform is also opening to all Cloudflare accounts for free.

The stated goal is a system that learns from every attempt.