tooluniverse-dose-response

Pass

Audited by Gen Agent Trust Hub on Oct 7, 2026

Risk Level: SAFE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests external data through the scripts/fit_dose_response.py script, which reads a CSV file provided by the user. The attack surface is effectively mitigated because the script strictly validates input by attempting to cast concentration and response values to floats. Any non-numeric data (including potential prompt injection strings) will trigger a ValueError and be skipped, preventing the data from influencing the agent's logic or being interpolated into a prompt.
  • Ingestion points: scripts/fit_dose_response.py (reads --input CSV file).
  • Boundary markers: The script enforces a strict CSV schema.
  • Capability inventory: The script performs local mathematical calculations and prints results to stdout; it has no network, file-write, or shell execution capabilities.
  • Sanitization: Strong sanitization via mandatory float casting (float(row["concentration"])).
  • [EXTERNAL_DOWNLOADS]: The skill depends on numpy and scipy. These are well-known, established scientific libraries widely used in the industry for data analysis. Their use is appropriate for the skill's primary purpose and does not represent a security risk.
  • [DATA_EXPOSURE]: The skill reads from a user-specified file path for its analysis. This is a standard requirement for a data fitting tool. There is no evidence of hardcoded credentials, access to sensitive system directories (like .ssh or .aws), or exfiltration of the data to external endpoints.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 7, 2026, 01:49 PM