tool-design

Pass

Audited by Gen Agent Trust Hub on Apr 7, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill advocates for 'Architectural Reduction', a design pattern where specialized tools are replaced by a single tool allowing arbitrary bash command execution (e.g., using standard Unix utilities like grep, cat, ls). While presented as a design principle for sandboxed environments, it encourages a pattern that increases the attack surface if implemented without rigorous security controls.
  • [PROMPT_INJECTION]: The 'Tool-Testing Agent Pattern' described in SKILL.md demonstrates a method for optimizing tool descriptions by feeding untrusted 'failure examples' back into an agent prompt.
  • Ingestion points: The failure_examples parameter in the optimize_tool_description function (SKILL.md).
  • Boundary markers: The prompt template lacks delimiters or instructions to ignore potential commands embedded within the failure data.
  • Capability inventory: The generated prompt is passed to an LLM via get_agent_response (SKILL.md).
  • Sanitization: The provided pattern does not include any escaping or validation of the input data before interpolation into the prompt.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 7, 2026, 07:35 PM