interview-system-designer

Warn

Audited by Gen Agent Trust Hub on Nov 12, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The script scripts/question_bank_generator.py uses dynamic module loading to import logic from its sibling file, scenario_stems.py. * Evidence: The script implements a _sibling function utilizing importlib.util.spec_from_file_location and spec.loader.exec_module to load code from a path computed at runtime using __file__.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process interview results and role definitions from JSON files, which creates a potential surface for indirect prompt injection. * Ingestion points: hiring_calibrator.py, loop_designer.py, and question_bank_generator.py all accept untrusted JSON data via the --input command-line argument. * Boundary markers: There are no explicit delimiters or instructions provided to the agent to distinguish between data and potential commands within these input files. * Capability inventory: The agent uses the output of these scripts to generate interview questions, rubrics, and calibration reports, allowing malicious data to influence agent behavior. * Sanitization: While the scripts perform structural validation of the JSON (e.g., checking for required fields and numeric ranges for scores), they do not sanitize or filter the string content for natural language instructions.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 12, 2026, 07:06 PM