marimo-pair
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses the
execute_codetool to run Python snippets in the user's active marimo kernel to inspect state and test transformations. - [DYNAMIC_EXECUTION]: The skill leverages the
marimo._code_modeinternal API to programmatically create, edit, and execute notebook cells and manage user interface elements at runtime. - [EXTERNAL_DOWNLOADS]: The skill provides instructions for managing Python packages via
ctx.packages.add()and fetching JavaScript libraries from the well-knownesm.shCDN for custom widgets. - [INDIRECT_PROMPT_INJECTION]: The skill processes data from the live notebook environment, creating an attack surface where content in notebook cells or variables could influence agent behavior.
- Ingestion points: Reading cell code via
ctx.cells[...].codeand inspecting live notebook variables. - Boundary markers: No explicit delimiters or 'ignore' instructions are provided to separate user-sourced notebook data from the agent's control logic.
- Capability inventory: Full Python code execution in the kernel via
execute_code, notebook structural modification viamarimo._code_mode, and package installation capabilities. - Sanitization: There is no evidence of sanitization or validation of content retrieved from the notebook before it is used to inform agent decisions.
Audit Metadata