explore-run

Warn

Audited by Gen Agent Trust Hub on May 18, 2026

Risk Level: MEDIUMREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The file scripts/write_outputs.py dynamically loads a Python module at runtime. It uses importlib.util to load and execute a script located at ../../../shared/scripts/write_explore_bundle.py. This pattern of loading code from a computed relative path is a security risk as it allows for the execution of arbitrary code if the target directory or file is manipulated.\n- [INDIRECT_PROMPT_INJECTION]: The skill exposes an attack surface for indirect prompt injection via data ingestion.\n
  • Ingestion points: The scripts/plan_variants.py script reads a JSON specification file through the --spec-json command-line argument.\n
  • Boundary markers: The script does not implement specific delimiters or 'ignore' instructions to isolate the ingested JSON data from the agent's logic.\n
  • Capability inventory: The skill has the capability to rank variants and prepare output files, and it hands off execution to other scripts which may perform training or system commands.\n
  • Sanitization: While the script performs basic data type conversion (e.g., safe_float), it lacks robust schema validation or sanitization of the input JSON contents.
Audit Metadata
Risk Level
MEDIUM
Analyzed
May 18, 2026, 10:23 AM
Security Audit — agent-trust-hub — explore-run