genie-metadata-optimizer

Pass

Audited by Gen Agent Trust Hub on Mar 8, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill installs the gepa package (version >=0.1.0) and other libraries such as mlflow and databricks-sdk from PyPI. These are utilized for the core GEPA (Genie Evolution through Prompt Adaptation) optimization logic and for interacting with Databricks services.- [COMMAND_EXECUTION]: The notebook template (run_gepa_optimization.py) executes %pip install and dbutils.library.restartPython() to manage its runtime environment. The skill also executes dynamic SQL statements via spark.sql() and the Databricks SQL Statement API to evaluate and score agent-generated responses.- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection. It extracts information from judge rationales and evaluation metadata, which are LLM-generated, and uses this data to update Genie Space instructions and metadata without sanitization.
  • Ingestion points: Failure rationales and counterfactual fixes are extracted from the eval_results input in metadata_optimizer.py.
  • Boundary markers: No delimiters or safety instructions are used when incorporating feedback into proposed patches.
  • Capability inventory: The skill possesses the ability to modify Genie Space configurations via the PATCH API and to execute arbitrary SQL queries for evaluation.
  • Sanitization: The skill does not perform validation or escaping of the ingested feedback strings before they influence configuration changes.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 8, 2026, 02:33 AM