Skip to content

AAK-POISON-002: Prompt injection patterns in tool description

Severity
CRITICAL
Category
Tool Poisoning (tool-poisoning)
Scanner
agent_audit_kit/scanners/tool_poisoning.py
Auto-fixable
No.
SARIF name
ToolDescPromptInjection

What it detects

Tool description contains prompt injection patterns such as 'ignore previous instructions', 'system:', or role-switching directives.

How to fix

Delete or reword the phrase quoted in the finding so the description only says what the tool does. The rule matches ignore previous, system:, you are, forget, new instructions, override, admin mode, <hidden> and <system> (case-insensitive), so a harmless phrase such as 'the files you are allowed to read' needs rewording too.

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

agent-audit-kit scan . --rules AAK-POISON-002

Limitations

Matches a fixed list of phrases: ignore previous, system:, you are, forget, new instructions, override, admin mode, <hidden> and <system>. The same instruction in other words is not matched, for example ignore all previous instructions (a word sits between the two), disregard the instructions above, or an <IMPORTANT> block that tells the model to keep a step from the user.

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-POISON-002.md: the page then shows that text, with the registry reference appended under it.