Third-party cyber evaluations involving OpenAI models
openai.com·15h ago
Security
Cloudflare Blog·May 19·3 min read
Announcing Claude Managed Agents on Cloudflare
TL;DR
Developers previously lacked integrated tools to run AI agents securely at scale with visibility into their operations. Cloudflare and Anthropic integrated Claude Managed Agents with Cloudflare Sandboxes—isolated execution environments—alongside new capabilities including an Agents SDK framework, Browser Run for programmable browser automation, and Dynamic Workers for scalable code execution.
✦ Why It Matters
Engineers can now deploy AI agents securely at scale with built-in monitoring and private service access.
Key Takeaways
- Use Agents SDK to build customizable AI agent frameworks without building infrastructure from scratch.
- Leverage Browser Run to give agents programmable web browser capabilities for automated interactions.
- Deploy code at scale using Dynamic Workers with automatic sandboxing for security isolation.
- Monitor agent behavior and secure private service connections through integrated observability tools.
Full Summary
Cloudflare and Anthropic have partnered to integrate Claude Managed Agents with Cloudflare Sandboxes, which are isolated runtime environments that execute code safely without affecting other systems. The collaboration addresses the challenge of deploying AI agents reliably at scale while maintaining security and observability—the ability to monitor and understand system behavior.
The solution includes three core components: Agents SDK, a customizable framework for building agents; Browser Run, which provides agents with fully programmable and observable web browsers for interaction; and Dynamic Workers, enabling sandboxed code execution across distributed infrastructure. This integration gives developers greater control over agent sandbox configurations, secures connections to private backend services, and improves visibility into agent operations.
The platform enables developers to deploy agents at massive scale while maintaining security boundaries and operational transparency.