Tool, function and MCP abuse
The agent is led to call its tools in harmful ways, or the tools themselves carry the attack: an MCP server whose tool descriptions contain hidden instructions, or one that changes its descriptions after it was approved.
- LLM03:2026 Excessive Agency
- ASI02:2026 Tool Misuse and Exploitation
- AML.T0053 AI Agent Tool Invocation
- AML.T0110 AI Agent Tool Poisoning
- AML.T0101 Data Destruction via AI Agent Tool Invocation
How it works
- 1Agents choose which tool to call, and with what arguments, from text: the user's request, the conversation and each tool's own description.
- 2A tool description is read by the model like any other text, so a malicious or compromised server can steer the agent through it.
- 3A server approved once can change its tools later. Unless someone pins and checks them, the agent uses the new version.
Real cases on the map
- CriticalPraisonAI: AICoder Arbitrary File Write and Command Execution via LLM Tool Callspraisonai
- HighvLLM has remote code execution vulnerability in the tool call parser for Qwen3-Codervllm
- MediumMCP Atlassian: Path traversal in upload_attachment allows arbitrary file read and exfiltration via MCP tool callmcp-atlassian
- MediumOpenClaw ACP client has permission auto-approval bypass via untrusted tool metadataopenclaw
Published advisories whose own titles describe this technique.
What to watch for
- Tool calls unrelated to the task
- Arguments reaching new paths, hosts or accounts
- A server's tool list or descriptions changing between sessions
How to defend
- Allow only the tools a task needs; scope each tool's credentials.
- Pin MCP server versions and review description changes before accepting them.
- Ask a person before destructive or outbound actions.