Workflow
Audit MCP Tool Security
A checklist that catches the failure modes unique to model-driven tool calls.
1 min readupdated 2026-08-01
/ quick answer
Treat every tool call as untrusted input and every model output as untrusted intent. A checklist that catches the failure modes unique to model-driven tool calls.
A checklist that catches the failure modes unique to model-driven tool calls. The problem it solves: An MCP tool is callable by a model that can be manipulated through the content it reads. Treat every tool call as untrusted input and every model output as untrusted intent. It runs in 7 steps, starting with assume prompt injection: content the agent reads can try to trigger your tools. This workflow node is part of the Onexial knowledge graph and links to related concepts, workflows and tools below.
Problem
An MCP tool is callable by a model that can be manipulated through the content it reads.
Solution
Treat every tool call as untrusted input and every model output as untrusted intent.
Steps
- 01Assume prompt injection: content the agent reads can try to trigger your tools.
- 02Remove generic escape hatches (`run_sql`, `exec`, `http_request`) from the surface.
- 03Enforce authorisation inside the server, never in the tool description.
- 04Cap output size and redact secrets, tokens and PII before returning.
- 05Require explicit approval for destructive and money-moving tools.
- 06Rate limit per user and alert on unusual call patterns.
- 07Re-test after every tool addition — new tools create new combinations.
Tools Used
Related Dictionary
/ frequently asked
What does the Audit MCP Tool Security workflow do?
Treat every tool call as untrusted input and every model output as untrusted intent.
What problem does Audit MCP Tool Security solve?
An MCP tool is callable by a model that can be manipulated through the content it reads.
How many steps does Audit MCP Tool Security take?
7 steps. It starts with assume prompt injection: content the agent reads can try to trigger your tools. and ends with re-test after every tool addition — new tools create new combinations..
Which tools does Audit MCP Tool Security need?
It uses mcp-integration-stack, ai-security-stack — each linked below with its own node.
↳ connected nodes
Dictionary↳ linked
MCP Tools
MCP tools are typed, described functions an AI model can call — the unit of capability that decides whether an agent is useful or dangerous.
Dictionary↳ linked
MCP Transport (stdio vs Streamable HTTP)
MCP runs either as a local stdio process or as a remote streamable HTTP endpoint — the choice decides auth, deployment and who can reach it.
Dictionary↳ linked
Prompt Injection
An attack where hostile input hijacks the LLM's instructions, causing it to leak data or misbehave.
Tool Stack↳ linked
MCP Integration Stack
Build, deploy and secure MCP servers that real AI clients can use.
Tool Stack↳ linked
AI Security Stack
Least-privilege tooling, approval gates and audit trails for agentic systems.
Dictionary↳ linked
AI Security
AI security protects systems where the model is an untrusted decision-maker acting on untrusted input with real tool access.