AI Agent and MCP Security: Threats, Controls and a Practical Checklist

ASA AI Team · Updated 2026-11-08 · 4 min read

AI agent security is the discipline of controlling what autonomous AI systems can see and do. With the Model Context Protocol (MCP), agents connect to tools and data through a standard interface — which makes them far more useful, and turns every connected tool into part of the attack surface.

Key takeaways

Why agents change the threat model

A chatbot can only produce text. An agent can read your mailbox, query databases, open tickets, commit code or move money. When it is manipulated — usually through prompt injection hidden in content it reads — the consequence is an action, not just a bad answer. Securing agents therefore means securing the tools and identities they use.

MCP in one paragraph

The Model Context Protocol is an open protocol that lets AI applications discover and call tools, read resources and use prompts exposed by MCP servers. A single agent can connect to many servers — internal ones you build and third-party ones you install. Each server describes its tools in natural language that the model reads and trusts.

Main threats to agents and MCP servers

ThreatHow it worksControl
Indirect prompt injectionInstructions hidden in documents, web pages or tool outputs steer the agent.Treat tool output as untrusted; scan content; separate data from instructions; require approval for sensitive actions.
Tool poisoningA tool description contains hidden instructions that the model follows.Review and pin tool definitions; scan descriptions; alert on any change.
Rug pullA trusted server silently changes a tool's behaviour or description after approval.Hash and version manifests; re-approve on change; prefer internally hosted servers.
Tool shadowingOne server's tool descriptions influence how the agent uses another server's tools.Isolate servers per agent and task; limit which servers can be combined.
Excessive permissionsTools run with broad, long-lived credentials.Least privilege; per-tool and per-user scopes; short-lived tokens.
Token misuseCredentials are passed through or reused across services.Validate audience and scope of every token; no token pass-through.
Data exfiltrationThe agent leaks data through links, images, e-mails or file writes.Egress allow-lists; output scanning; block rendering of untrusted links and images.
Shadow agentsTeams deploy agents and MCP servers nobody has reviewed.Discovery and inventory of agents, servers, tools and owners.

A practical checklist

  1. Discover every agent, MCP server and tool, with an owner for each.
  2. Classify tools by impact: read-only, reversible write, irreversible or external action.
  3. Restrict each agent to an allow-list of servers and tools for its role.
  4. Scope credentials per tool and per user, short-lived, never embedded in prompts.
  5. Require human approval for irreversible or high-impact actions.
  6. Pin and monitor tool definitions; alert on any change.
  7. Filter inputs and outputs with runtime guardrails.
  8. Control egress so agents can only reach approved destinations.
  9. Log every tool call with user, agent, parameters and result, and send it to your SOC.
  10. Red team the agent regularly, including indirect injection through each tool.

Where to enforce controls

Controls that live only in the system prompt can be talked around. Effective enforcement sits in an MCP gateway or proxy between agents and servers, in the identity layer that issues scoped tokens, and in runtime monitoring that sees every call. This is the approach we take with SENTINEL: discovery of agents and MCP servers, policy enforcement on every tool call, and detection mapped to OWASP and MITRE ATLAS.

Frequently asked questions

What is MCP security?

MCP security covers the risks introduced when AI agents connect to tools and data through the Model Context Protocol: tool poisoning, changed tool definitions, over-scoped credentials, cross-server interference and data exfiltration, and the controls that address them.

What is tool poisoning?

Tool poisoning is an attack where malicious instructions are hidden in a tool's description or metadata. Because the model reads these descriptions to decide how to use tools, it may follow the hidden instructions.

Can a system prompt protect an AI agent?

Not on its own. Prompts can be bypassed through injection. Permissions, approvals, egress control and monitoring must be enforced outside the model.

How do I start securing AI agents?

Start with an inventory of agents, MCP servers and tools, apply least privilege and human approval for sensitive actions, then add runtime monitoring and regular red teaming.

Related guides