Darktrace Warns of AI Agent Risks in Model Context Protocol

Darktrace’s blog on MCP risks explains that Model Context Protocol has become the control plane for autonomous AI agents, expanding the attack surface across development pipelines, operational systems and customer workflows. It notes that MCP is a standard introduced by Anthropic and that risky configurations—such as overly permissive agents and prompt injections—can lead to data exfiltration, unwanted actions, or compromised workflows if governance and visibility are lacking.

The piece outlines seven key MCP risks, including content-injection adversaries, tool abuse and over-privileged agents, cross-agent contamination, supply chain risk, unintentional agent behaviours, confused deputy attacks, and governance blind spots.

To counter these threats, it recommends treating MCP as critical infrastructure, enforcing identity and least-privilege controls, validating inputs and outputs, deploying sandboxed testing environments, and implementing provenance tracking, cryptographic signing, and CI/CD security gates for MCP integrations. It also emphasises ongoing monitoring, incident planning, and governance policies, such as ISO 42001, to ensure proper ownership, auditing and response for MCP deployments.

Written by Shanita Sojan, the analysis underscores the need for governance, visibility and safeguards before MCP-driven automation becomes deeply embedded.