AI Agent Tool and MCP Security

What is AI agent tool security?

AI agents act on the world through tools: connectors that let them search logs, query a ticketing system, isolate a device or send a message. AI agent tool security is the work of making sure those tools are trustworthy, carry only the permissions they need and cannot be turned against the organization.

The Model Context Protocol (MCP) is an open standard for connecting AI applications to tools and data sources. It has become a common way to give agents tools, which makes MCP servers an important part of the attack surface.

What are the main risks?

Malicious or impersonated tool servers that steal data or return poisoned results. Anomali Cyber Watch has reported on malicious MCP servers.
Tool poisoning: hidden instructions in a tool's description or output that manipulate the agent. See prompt injection.
Over-broad permissions: a tool holding admin credentials when the agent only needs read access.
Classic vulnerabilities in tool code, such as command injection or unrestricted URL fetching.
Weak identity: unclear which agent, acting for which user, made a request. See non-human identity.

How do you secure agent tools?

Keep an approved list of tools and servers, and block everything else.
Give each tool least-privilege credentials scoped to the agent's job.
Treat tool output as untrusted data, not instructions.
Separate read tools from action tools, and put action tools behind approval gates.
Log every tool call with the agent, the user it acted for and the result.
Review and patch tool code like any other software.

Frequently asked questions

Is MCP itself insecure?

MCP is a protocol. The risks come mostly from how individual servers are built, what permissions they hold and whether agents trust them blindly.

Who should own MCP security?

Usually the team that owns the agents, with application security reviewing tool code and identity teams managing the credentials tools use.

Anomali and AI Agent Tool and MCP Security

Anomali Threat Research tracks threats to AI tooling, including malicious MCP servers, in Anomali Cyber Watch.

Request a demo to see how Anomali governs its own agents.