meltpool-geometry

Warn

Audited by Gen Agent Trust Hub on Nov 7, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill fetches regression and classification CSV datasets from the author's GitHub repository. These files are stored locally in a cache directory (~/.cache/mechanical-engineering-skills/meltpoolnet) to support the skill's machine learning functionality.
  • [DYNAMIC_EXECUTION]: The predict tool uses joblib.load() to deserialize model files from paths provided by the user. Because joblib uses pickle internally, it can execute arbitrary Python code during the loading process if a malicious model file is provided. This is a common pattern in machine learning tools but requires the user to trust the source of the model files.
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources which is then processed and returned to the agent.
  • Ingestion points: scripts/meltpool.py reads data from downloaded GitHub CSVs and user-supplied CSV files via the --extra argument.
  • Boundary markers: None present in the output JSON provided to the agent.
  • Capability inventory: The script can read/write files in the cache and user-specified paths, and it outputs data used by the agent for reporting.
  • Sanitization: Data undergoes numeric coercion via pandas, but no content validation is performed to prevent malicious strings in non-numeric fields from reaching the agent.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 7, 2026, 06:02 AM