kanban-agent
Warn
Audited by Gen Agent Trust Hub on Nov 8, 2026
Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data (task titles, descriptions, checklist items) from a remote Kanban board API which is then provided to the agent.
- Ingestion points: Data is retrieved via
kb_list_tasks,kb_get_task, andkb_summaryinexports.py. - Boundary markers: No specific delimiters or "ignore embedded instructions" warnings are used when presenting board data to the agent in
SKILL.md. - Capability inventory: The skill has extensive write capabilities to the board (create/delete boards, columns, and tasks) and file system write capabilities via
kb_export_for_sharing. - Sanitization: No sanitization or filtering of external content is performed in
exports.pybefore returning data to the agent. - [DYNAMIC_EXECUTION]: The skill uses Python's
shutilandosmodules to programmatically modify and distribute its own source code. - Evidence: The
kb_export_for_sharingfunction inexports.pyreads its own source file, performs string replacement to inject a board URL into theDISTRIBUTED_BOARD_URLvariable, and writes the modified script to a new location on the file system (output/kanban-agent-share/). - Impact: While intended for legitimate sharing, this pattern involves generating executable code at runtime, which is a behavior that could be abused if the input URL is not properly validated.
Audit Metadata