---

⚠ Untrusted MCP Servers

Risk Level: 🔴 Critical
Affected: MCP Clients & AI Agents
How: AI assistants connected to untrusted MCP servers may receive malicious tools or manipulated data capable of influencing AI behaviour.
Impact: Enterprise data leakage | Credential exposure | Supply-chain compromise
Defence: Connect only to trusted MCP servers, verify server identity, restrict available tools, and continuously monitor server activity.
Reference: Model Context Protocol Specification

---

⚠ Excessive Tool Permissions

Risk Level: 🔴 Critical
Affected: Enterprise AI Agents
How: AI agents granted unrestricted access to email, cloud storage, source code, databases, or browsers can perform unintended actions if compromised or manipulated.
Impact: Unauthorised actions | Data exposure | Business process abuse
Defence: Apply least privilege, require user approval for high-risk actions, and review permissions regularly.
Reference: Enterprise AI Security Best Practices

---

⚠ Prompt Injection Through MCP Integrations

Risk Level: 🟠 High
Affected: MCP-enabled Applications
How: Hidden instructions embedded within emails, documents, web pages, or retrieved content manipulate AI behaviour through connected MCP tools.
Impact: Incorrect AI actions | Information disclosure | Automation abuse
Defence: Treat external content as untrusted, validate retrieved information, isolate high-risk workflows, and implement user confirmation for sensitive actions.
Reference: OWASP Top 10 for LLM Applications

---

⚠ Malicious MCP Tool Supply Chain

Risk Level: 🟠 High
Affected: MCP Tool Ecosystem
How: Compromised or malicious MCP tools may return manipulated responses, execute unauthorised operations, or exfiltrate enterprise information.
Impact: Supply-chain compromise | Persistent access | Sensitive data theft
Defence: Install MCP tools only from trusted publishers, verify updates, digitally sign internal tools, and continuously monitor tool activity.
Reference: Software Supply Chain Security Guidance

---

⚠ Identity and Secret Management

Risk Level: 🟠 High
Affected: MCP Servers & Connected Services
How: Poor management of API keys, OAuth tokens, and service credentials increases the risk of unauthorised access to connected enterprise systems.
Impact: Account compromise | Lateral movement | Cloud service abuse
Defence: Store secrets securely, rotate credentials regularly, implement short-lived tokens, and monitor authentication activity.
Reference: Identity Security Best Practices

---

⚠ Enterprise Governance for MCP

Risk Level: 🟡 Advisory
Affected: Enterprise AI Deployments
How: Rapid adoption of MCP without governance may introduce unmanaged AI integrations and excessive trust between AI agents and enterprise services.
Impact: Compliance risks | Operational disruption | Increased attack surface
Defence: Maintain an inventory of MCP servers and tools, implement Zero Trust principles, log AI activity, and conduct regular security assessments.
Reference: NIST AI RMF | Enterprise AI Governance

---

Shilpa Sayura Cyber Defence Research Lab

Editorial Note: MCP Security Watch is a special edition of Cyber Defence Digest. It provides practical guidance for organisations deploying AI agents and Model Context Protocol integrations. As enterprise AI adoption accelerates, securing MCP ecosystems should become part of every organisation’s identity, application, and supply-chain security strategy.