Anthropic adds MCP tunnels and self-hosted sandboxes to Claude Managed Agents for enterprise security
Anthropic has added two enterprise security features to Claude Managed Agents: MCP tunnels, which route agent services through private networks without public internet exposure, and self-hosted sandboxes, which keep sensitive tool execution within customer infrastructure while Anthropic handles orchestration.
Anthropic adds MCP tunnels and self-hosted sandboxes to Claude Managed Agents for enterprise security
Anthropic has launched two new security features for Claude Managed Agents that allow enterprises to keep agent operations within their own network perimeters.
MCP tunnels enable private network routing
The MCP tunnels feature allows Claude Managed Agents to access Model Context Protocol (MCP) servers inside private networks without exposing them to the public internet. According to Anthropic, internal databases, private APIs, knowledge bases, and ticketing systems can become agent-accessible tools through a lightweight gateway that makes a single outbound connection with no inbound firewall rules required.
Traffic is encrypted end-to-end, and no public endpoints are created. MCP tunnels is currently available as a limited research preview requiring access approval.
Self-hosted sandboxes move tool execution on-premises
The self-hosted sandbox feature separates agent orchestration from tool execution. Anthropic's infrastructure continues to handle the agent loop, context management, and error recovery, while tool execution moves to customer-controlled environments or managed sandbox providers.
Sensitive files, packages, and services remain in customer infrastructure. Anthropic has partnered with Cloudflare, Daytona, Modal, and Vercel for managed sandbox options, though customers can also bring their own sandbox client. Self-hosted sandboxes launched as a public beta feature.
"Both the sandbox where an agent executes tools and the services it reaches run within the established boundaries of your enterprise, under your security and runtime controls," Anthropic stated.
Context: Recent Managed Agents updates
Anthropic launched Claude Managed Agents in April 2025 as a simplified way to build and deploy cloud-hosted AI agents. Earlier in May, the company added dreaming, outcomes, and multiagent orchestration features.
Karpathy joins Anthropic
Separately, OpenAI founding team member and research scientist Andrej Karpathy announced he has joined Anthropic. "I think the next few years at the frontier of LLMs will be especially formative," Karpathy stated on social media.
What this means
These features address a critical enterprise barrier to AI agent adoption: data security and network isolation. By allowing agents to operate entirely within private networks while maintaining Anthropic's orchestration layer, companies can deploy Claude agents without creating new security vulnerabilities. The self-hosted sandbox approach is particularly significant—it lets enterprises maintain zero-trust architectures while still leveraging cloud AI capabilities. Both features suggest Anthropic is prioritizing enterprise deployment over pure technical advancement, focusing on making existing capabilities deployable at scale rather than pushing frontier model performance.
Related Articles
Google Announces Gemini Spark Agent and Antigravity Platform at I/O, Launch Date Not Disclosed
Google announced Gemini Spark at I/O 2026, positioning it as a competitor to OpenAI's Claude-based agents. The service will integrate with Gmail, Calendar, Drive, and other Google apps, running on Gemini 3.5 Flash and a new platform called Antigravity. No general availability date has been disclosed.
Google opens CodeMender API to select testers, pitching AI security tool to governments and enterprises
Google announced at I/O 2026 that it is opening API access for CodeMender, its AI agent for code security, to select expert groups. The company is positioning the tool to compete with Anthropic's Mythos Preview, which flagged unknown security vulnerabilities and secured major government and enterprise contracts.
AWS releases four multimodal evaluators for image-to-text AI tasks in Strands Evals SDK
AWS has added four multimodal evaluators to its Strands Evals SDK that judge image-to-text AI outputs by directly analyzing source images. The evaluators—Overall Quality, Correctness, Faithfulness, and Instruction Following—use multimodal large language models to detect visual hallucinations, factual errors, and instruction violations that text-only judges miss.
AWS SageMaker AI adds bidirectional streaming for real-time speech transcription with vLLM
Amazon SageMaker AI has launched bidirectional streaming support for real-time inference, enabling WebSocket-based voice applications through vLLM integration. The feature uses HTTP/2 on port 8443 to bridge client connections with vLLM's Realtime API, allowing audio to stream in while transcription streams back simultaneously over a single persistent connection.
Comments
Loading...