Skip to content

AAK-LLM-001: Model flagged tool metadata as manipulative

Severity
HIGH
Category
Tool Poisoning (tool-poisoning)
Scanner
No scanner module. The rule id is named in code by agent_audit_kit/llm_scan.py.
Auto-fixable
No.
SARIF name
LlmFlaggedToolMetadata

What it detects

With scan --llm-scan, the chosen model read a tool's metadata (an MCP tool's name, description and parameter descriptions, or a skill's instructions) and judged that it tries to manipulate the agent that reads it: an action the user did not ask for, collecting or sending data the tool does not need, hijacking other tools, overriding the agent's instructions, impersonation, deception or concealment. This is a model's judgement, not a pattern match; the evidence quotes its reason.

How to fix

Read the flagged text and the model's reason. Remove any instruction aimed at the agent rather than the user: hidden directives, requests to read or send data the tool does not need, references that steer other tools. If you do not control the tool, stop using it until its publisher fixes the description.

To confirm the finding is gone, report only this rule:

agent-audit-kit scan . --rules AAK-LLM-001

Limitations

Reads only metadata AAK can extract without running anything: inline tools in MCP configs, saved tools/list responses (*tools*.json), Python tools registered with @x.tool, @tool or Tool(name=..., description=...) string literals, TypeScript/JavaScript server.tool(...) and registerTool(...) string literals, and SKILL.md. A description a server builds at runtime, concatenates from several strings, or returns only from a live tools/list is never seen. The verdict is a model's: it can miss manipulation phrased as ordinary documentation and flag benign text.

References


All rules. This page is generated from the rule registry when the documentation is built; to change it, change the rule in agent_audit_kit/rules/builtin.py. To give the rule a hand-written write-up, add docs/rules/AAK-LLM-001.md: the page then shows that text, with the registry reference appended under it.