Explore our new Cybersecurity Terms Glossary, explained by our experts.

View the Glossary View the Glossary
PRIVIA

Cybersecurity Terms Glossary

What Is Tool Poisoning?

Tool poisoning hides malicious instructions in the description of a tool an AI agent can call, so the payload reaches the model through the integration layer itself.

Illustration for the cybersecurity term Tool Poisoning
Home / Blog / Cybersecurity Terms / What Is Tool Poisoning?
17 November 2026

What Is Tool Poisoning?

What Is Tool Poisoning?

When an AI agent decides which tool to call, it reads a natural-language description of each one: what the tool does, what parameters it takes, when it should be used. Those descriptions come from whoever wrote the tool, they are placed directly into the model’s context, and the model has no way to distinguish them from instructions written by the application developer. Tool poisoning is the exploitation of that fact—a malicious description that tells the model to do something other than what the user asked.

Disclosure and Mechanics

Invariant Labs published the technique in April 2025 in the context of the Model Context Protocol, demonstrating a server whose innocuous-looking tool carried hidden instructions in its description: before performing the advertised function, read a configuration file and include its contents in the parameters. The user sees a tool called “add numbers” and an approval dialogue showing a plausible call. The model, having read the full description, behaves as instructed. Because interfaces typically show users a summary rather than the raw metadata, the payload is invisible at exactly the moment a human is asked to approve.

Variants

Why It Is Worse Than It Looks

Tool descriptions sit in a gap between two review processes. They are not code, so they escape code review. They are not user input, so they escape input validation. They arrive through an integration a developer added in an afternoon, and they are refreshed from the server at runtime, so what was reviewed at installation is not necessarily what the model reads today. Meanwhile the agent executing them usually holds broad credentials, which turns a successful poisoning into agent hijacking with whatever permissions were granted.

Controls

Pin and verify. Treat a connector like a dependency: a named version, a verified hash, an approved registry rather than an arbitrary repository, and re-review on update rather than silent acceptance. Review the raw tool metadata as untrusted text, looking for instructions aimed at the model rather than at a developer, and automate that check where you can. Surface the actual call to the approving user—tool name, destination, and full parameters—rather than the model’s description of what it intends. Isolate servers by sensitivity so that one connector reaching confidential data and another reaching the internet are not available in the same session. Scope credentials per tool and keep them short-lived, so a poisoned description reaches very little. Log every call with its arguments, which is the only way to notice a parameter quietly carrying a file path.

Governance in a UK Context

The first problem is usually discovery, because connectors are installed by individual developers rather than procured. Find what is running, assign an owner to each, and bring connectors into the approved software list and change control. Where a connector mediates access to personal data, it belongs in your record of processing and in access reviews, and an exfiltration through it is a personal data breach reportable to the ICO. Financial entities in scope of DORA should treat an agent connector reaching a critical system as part of the ICT estate rather than as developer tooling. OWASP covers the underlying problem through excessive agency and supply chain entries in its LLM Top 10, and our note on over-privileged AI agents covers the permission side. At Privia Security, we review connector inventories, tool metadata, and agent permissions through our AI Penetration Test Services.

FAQ – Frequently Asked Questions

How Is Tool Poisoning Different From Prompt Injection?

It is a delivery channel for it. Ordinary indirect injection hides instructions in content the model reads while working—a document or a web page. Tool poisoning places them in the integration layer’s own metadata, which users never see and reviewers rarely read, and which the model treats as authoritative because it describes its own capabilities.

Will an Approval Prompt Protect Us?

Only if it shows the real call. Many interfaces display a model-generated summary of what it intends to do, which the poisoned description can influence. Show the tool name, the destination, and the full parameter values, and make approval meaningful by reserving it for actions that matter rather than asking for it on every step.

How Do We Vet a Connector?

Read the raw tool metadata as untrusted text and look for instructions aimed at the model. Check the publisher and whether source is available, identify the credentials and scopes required, confirm it can run isolated, and pin a reviewed version with a verified hash. Repeat the metadata review on every update, because descriptions can change without any visible signal.

Does This Only Affect MCP?

No. Any architecture where a model reads natural-language descriptions of tools it can call is exposed, including function calling in provider APIs, plugin systems, and in-house agent frameworks. MCP is where the technique was first demonstrated publicly, largely because its ecosystem of community servers made the attack easy to illustrate.

Related Terms

You May Be Interested In These

What Is Zerologon?

Cybersecurity Terms

17 November 2026

What Is Zerologon?

Read More
What Is Tenant Isolation?

Cybersecurity Terms

17 November 2026

What Is Tenant Isolation?

Read More
What Is ISO/IEC 27005?

Cybersecurity Terms

17 November 2026

What Is ISO/IEC 27005?

Read More