check-deliverable

Pass

Audited by Gen Agent Trust Hub on Jul 9, 2026

Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: Argument interpolation: User-supplied values for --tier and --subgroup are placed directly into the instruction block for the sub-agent without explicit validation or sanitization, which could allow for minor instruction overriding if the platform doesn't enforce strict argument enumerations.\n- [DATA_EXFILTRATION]: File access: The skill enables the agent to access and process the contents of arbitrary local files based on a user-provided path, though this is restricted to text-based extensions (.md, .txt, .markdown) which mitigates the risk of harvesting standard system credentials.\n- [COMMAND_EXECUTION]: Sub-agent instructions: The skill directs a sub-agent to perform file-level operations, specifically to "Read and Grep the file directly". While this is not a direct shell execution by the primary agent, it defines a capability surface that relies on the sub-agent's underlying safety when handling file paths.\n- [PROMPT_INJECTION]: Indirect prompt injection risk:\n
  • Ingestion points: The agent ingests and processes the content of external text files provided by the user (SKILL.md, Step 2).\n
  • Boundary markers: The instructions provided to the sub-agent lack explicit boundary markers (e.g., XML tags or triple quotes) or specific directives to ignore instructions embedded within the target file.\n
  • Capability inventory: The skill uses the Agent tool to spawn a sub-agent (qa-reviewer) which is granted file-reading and analysis capabilities (SKILL.md, Step 2).\n
  • Sanitization: The skill performs file extension validation to restrict processing to text files, but does not sanitize the contents of the file before processing.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 9, 2026, 12:03 PM
Security Audit — agent-trust-hub — check-deliverable