draft-validation
Pass
Audited by Gen Agent Trust Hub on Nov 15, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONCREDENTIALS_UNSAFE
Full Analysis
- [DYNAMIC_EXECUTION]: The skill uses
exec()within theEditValidation.accept()method to process and run Python code generated or modified by an LLM. - Evidence:
SKILL.mdstates: "accept() runs the revised code with exec(). Read it with to_code() or review() first..." - Mitigations: The skill implements
validate_syntax()which checks if the code parses and ensures only permitted validation methods are used before execution. It also providesedit.review()andedit.diff()for human-in-the-loop verification. - [INDIRECT_PROMPT_INJECTION]: The skill processes data summaries and natural language instructions via LLMs to generate code, creating an attack surface for indirect prompt injection if the processed data contains malicious instructions.
- Ingestion points:
df(dataframes) processed viaDataScanand user-providedinstructionstrings inEditValidation. - Boundary markers: The skill notes that models only see a JSON summary of the table, not raw rows, reducing the injection surface.
- Capability inventory: Can perform file reads (YAML/Python paths) and execute code via
exec(). - Sanitization: Employs
validate_syntax()to restrict the generated code to knownpointblankvalidation methods. - [COMMAND_EXECUTION]: The skill documents terminal-based interactive sessions and environment variable management for API keys.
- Evidence: Documentation for
pb.assistant(display="terminal")andexportcommands for API keys inreferences/providers-reference.md. - [CREDENTIALS_UNSAFE]: The documentation shows examples of hardcoding API keys, though it explicitly recommends using environment variables as a safer alternative.
- Evidence:
SKILL.mdcontains a placeholder exampleapi_key="sk-...". The instructions correctly guide users to store secrets in.envfiles or environment variables.
Audit Metadata