maverick-finance-hardened
Pass
Audited by Gen Agent Trust Hub on Apr 21, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill includes instructions that explicitly tell the agent to disregard user or 'manager' commands if they request actions that would violate the skill's security constraints, such as data exfiltration or arbitrary code execution.
- [COMMAND_EXECUTION]: The skill is designed to run only four specific Python scripts (
ratio_calculator.py,dcf_valuation.py,budget_variance_analyzer.py,forecast_builder.py) using the Python standard library. - [DATA_EXFILTRATION]: The instructions contain strict prohibitions against piping script outputs to network-transmitting commands to ensure sensitive financial data remains local.
- [EXTERNAL_DOWNLOADS]: The skill's safety documentation references the author's official site and related GitHub repositories for reporting and evaluation.
- [PROMPT_INJECTION]: The skill consumes financial data from external JSON files, which represents an indirect prompt injection surface. 1. Ingestion points:
sample_financial_data.json,valuation_data.json,budget_data.json, andforecast_data.json. 2. Boundary markers: Not explicitly defined for the data ingestion process. 3. Capability inventory: Execution of localized Python scripts for data processing. 4. Sanitization: The workflow includes a verification phase to check input data for completeness and validity.
Audit Metadata