Tool poisoning

Compromised tools,contained at the gateway.

A tool server can hide instructions in its descriptions and results. Zenstra’s gateway checks every MCP and API call against the agent’s grant, inside your cloud.

The risk

Agents trust
their tools.

An agent follows what its tools tell it. Poison one tool and every agent using it takes orders from someone else.

  • Poisoned descriptionsInstructions hidden in how a tool describes itself.
  • Rogue serversAn MCP server added without review, or quietly replaced.
  • Over-scoped toolsA tool that can transfer when the task needed a read.
A real pathA trusted tool server changes overnight.
  1. 1mcp serverA server your agents already use publishes a new version.
  2. 2tool descriptionThe description now tells agents to include API keys in every request.
  3. 3gatewayThe change is flagged and the instruction quarantined, inside your cloud.
  4. HoldThe new version is held for review before any agent can call it.mcp kb-sync · v2.4 · awaiting review
Industry references
  • ASI04Agentic Supply Chain VulnerabilitiesOWASP Top 10 for Agentic Apps
  • AML.T0110AI Agent Tool PoisoningMITRE ATLAS
  • AML.T0109AI Supply Chain Rug PullMITRE ATLAS
Policy as code

How it is
stopped.

Agents trust their tools. Zenstra does not: every tool is reviewed, scoped as a grant and checked on every call.

  • Decided inline
  • Fails closed
  • Versioned with every decision
holdmcp.serverwhen version changesdenytool.descriptionwhen it carries instructions · quarantinedenymcp.callwhen tool not in grantallowkb.searchfrom approved servers
How Zenstra handles it

Every tool call
checked against a grant.

The gateway in your VPC is the decision point between agents and the tools they use.

Inventory

Know every tool server

MCP servers in use are inventoried, and new ones are flagged for review.

  • Every MCP server agents call
  • Unapproved servers flagged
  • Coverage shown per platform
Scope

Tools as scoped grants

Each tool is a grant with a scope and an expiry. A call outside it is blocked, whatever the tool says.

  • Scope per tool and per agent
  • Per-call and per-session expiry
  • Signed writes to memory
Contain

Stopped inside your cloud

Tool content carrying instructions is quarantined and the call is blocked before data moves.

  • Gateway in your VPC or Kubernetes cluster
  • Payloads never leave your environment
  • Every block on the record
Attack scenarios

How tools
get turned.

  • Description injection

    A tool’s description tells agents to do more than they should.

  • Result injection

    Data returned with instructions mixed in.

  • Server swap

    A trusted server replaced by one that behaves differently.

  • Scope creep

    A tool that gains write or transfer abilities over time.

  • Unreviewed servers

    MCP servers added to coding agents on a laptop.

  • Memory writes

    Tools writing unsigned content into shared memory.

FAQ

Common questions.

What is tool poisoning?

An attack where a tool an agent uses, such as an MCP server, carries instructions or behaviour that turns the agent against its task.

Where is it stopped?

At the Zenstra gateway in your cloud, and on devices by Zenstra AI Discovery for MCP servers added to local coding agents.

Do we need to vet every MCP server ourselves?

Zenstra inventories them and checks every call against a grant, so a server you have not vetted cannot reach beyond what the task allows.