tool-design

Warn

Audited by Gen Agent Trust Hub on Apr 7, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill documentation (SKILL.md and references/architectural_reduction.md) advocates for providing agents with tools for arbitrary bash command and SQL execution. This 'Architectural Reduction' pattern grants agents high-privilege system access, which can be misused if the sandbox environment is improperly configured.
  • [PROMPT_INJECTION]: The skill describes a 'Tool-Testing Agent Pattern' that ingests external data (tool specifications and failure examples) into agent prompts for optimization.
  • Ingestion points: Untrusted data enters the agent context in the optimize_tool_description function in SKILL.md.
  • Boundary markers: The provided prompt template lacks clear delimiters or 'ignore' instructions for the embedded data.
  • Capability inventory: The skill encourages tools with broad capabilities, including bash and SQL execution.
  • Sanitization: No validation or escaping of the input data is described.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 7, 2026, 12:58 PM