The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external, untrusted sources (document titles, content outlines, and user comments) to drive its knowledge organization and permission governance workflows. This provides a vector for an attacker to influence the agent's actions through malicious content embedded in those documents.
Ingestion points: The agent ingests untrusted data via tools such as drive files list, drive metas batch_query, docs +fetch, and drive +list-comments as described in files like lark-drive-workflow-knowledge-organize-analysis.md and lark-drive-workflow-topic-move-collector-recall.md.
Boundary markers: The skill implements strong boundary markers through its state-machine-driven workflows (e.g., in lark-drive-workflow-knowledge-organize.md), which explicitly require user confirmation for the 'Approach Decision' and 'Move/Create Plan' before any destructive or high-risk operations are executed.
Capability inventory: The skill possesses powerful capabilities including drive +move, drive +delete, wiki +node-create, and drive +apply-permission, which could be abused if an injection were successful.
Sanitization: Data is normalized into structured models like ResourceItem and PlanItem during processing, which helps to separate raw content from the execution logic.

lark-drive