Anthropic Launches Security Overhaul for Claude Agents: Credentials No Longer Travel Inside AI

Anthropic Launches Security Overhaul for Claude Agents

Anthropic today released two new features for Claude Managed Agents that directly address the biggest barrier to enterprise AI agent adoption: the risk of credential leaks. The new self-hosted sandboxes and MCP tunnels ensure authentication tokens never pass through the agent itself, keeping sensitive keys within corporate infrastructure.

The announcement marks a critical shift in how enterprise AI agents are deployed. Instead of carrying credentials as part of the agent's execution context, the new architecture moves credential control to the network boundary. “This is a fundamental redesign of agent security,” said Dr. Alice Chen, product lead at Anthropic. “Enterprises can now trust that even if an agent is compromised, the keys to their internal systems remain locked inside their own network.”

Self-hosted sandboxes, available in public beta today, allow enterprises to run tool execution within their own infrastructure while the agent loop stays on Anthropic's platform. MCP tunnels, currently in research preview, establish secure, outbound-only connections to private MCP servers without exposing credentials in the agent's context.

New Security Architecture

The architectural distinction Anthropic is making is clear: split the agent loop from tool execution. The orchestration, context management, and error recovery happen on Anthropic's infrastructure. Tool execution and credential handling happen within the enterprise's own systems. “This separation doesn't exist in other sandbox approaches,” noted cybersecurity analyst Mark Rivera of Gartner. “OpenAI's local execution moves everything to the client, but it doesn't isolate credentials the way Anthropic's split architecture does.”

For teams already using Claude Managed Agents, the recommended starting point is self-hosted sandboxes. This allows organizations to move tool execution onto their own infrastructure and test the boundary first. MCP tunnels, being in research preview, should be examined after hardening the sandbox deployment.

Background: The Credential Problem

Until now, enterprises have been slow to connect AI agents to internal APIs and databases. The reason isn't model capability—it's security. In most production deployments, agents carry authentication tokens as they execute tool calls. A compromised or misbehaving agent then has everything it needs to cause damage. “The credentials were effectively inside the agent's brain,” Rivera explained. “That's a catastrophic risk for any organization with sensitive data.”

Anthropic's solution moves credentials out of the agent context and into the network perimeter. Self-hosted sandboxes keep files and packages within the enterprise's infrastructure. MCP tunnels provide private network connectivity through a lightweight, outbound-only gateway—no credentials pass through the agent.

What This Means for Enterprise AI Deployments

For orchestration teams, these capabilities represent more than a security update—they enable better agent performance. By decoupling tool execution from the agent loop, enterprises gain granular control over compute resources, data access, and workflow mapping. “You can now design agents that handle sensitive operations without ever holding the keys,” said Chen. “That's a game-changer for regulated industries like finance and healthcare.”

The practical path forward is clear: start with sandboxes to move tool execution in-house, then explore MCP tunnels for private connectivity. Teams evaluating Claude Managed Agents for the first time should prioritize this split architecture as a core requirement. With competitors like OpenAI adding local execution options, the industry is recognizing that agent security demands architectural change—not just patchwork fixes.

Anthropic's release signals that the era of trusting AI agents with master keys is ending. “The future of enterprise AI is not about making agents smarter—it's about making them safer,” Rivera concluded. “Anthropic just took a huge step in that direction.”

Tags:

Anthropic Launches Security Overhaul for Claude Agents: Credentials No Longer Travel Inside AI