Cloudflare Launches Dynamic Sandboxes for AI Agents

Cloudflare’s new offering targets AI agents capable of performing tasks autonomously. Traditional cloud services typically require email verification, passwords and manual registration, which can slow automated deployments. The service allows agents to obtain temporary accounts and use resources such as Workers, lowering the operational barriers between a programmatic decision and live deployment.

Cloudflare has launched temporary accounts for AI agents with a default lifespan of 60 minutes. During that period, agents can claim an account and directly deploy Workers and other assets. Users who want to retain the resources created can convert the temporary account into a permanent one before it expires. The available information does not disclose the exact launch date, pricing or usage limits.

Cloudflare operates global networking and cybersecurity services backed by a vast, continuously updated codebase that traditional manual audits cannot comprehensively cover. The company has now disclosed an internal AI-assisted framework that introduces agents into the vulnerability discovery, validation and remediation process, helping engineering teams identify security risks that are genuinely exploitable.

Cloudflare's latest disclosure shows that AI agents first scan large codebases before different models cross-check potential vulnerabilities to filter out false positives and invalid findings. Items that pass the screening include a proof of concept and recommended fixes before being handed to engineering teams. The company did not disclose a specific date, the number of vulnerabilities or the amount invested.

Cloudflare services such as DNS and Workers span more than 2,500 API endpoints. Presenting each endpoint separately to an AI model would consume large numbers of tokens and raise inference costs. MCP is a standard protocol connecting AI agents to external tools, and the new server is designed to reduce integration complexity while preserving access controls and security boundaries.

Cloudflare’s newly released MCP-based server uses “Code Mode” to consolidate more than 2,500 API endpoints into two tool interfaces, allowing AI agents to dynamically write and execute code for each task. The code runs in an isolated V8 environment, while OAuth 2.1 strengthens authentication and permission controls for operations involving DNS and Workers.