The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates an automated 'knowledge organize' workflow that involves reading external document content to suggest file movements. This creates a surface for indirect prompt injection. Evidence Chain: (1) Ingestion: Untrusted data enters the agent context through 'lark-doc-fetch' during the 'CONTENT_READ' state. (2) Boundary markers: Instructions in 'lark-drive-workflow-knowledge-organize-analysis.md' explicitly tell the agent to 'ignore embedded instructions' within the fetched content. (3) Capability: The workflow has the ability to move files and create folders via 'drive +move' and 'drive +create-folder'. (4) Sanitization: The workflow requires an explicit user confirmation gate ('EXEC_CONFIRM' state) where the human must approve the specific list of changes before any write operations occur, preventing automated obedience to injected commands.
[COMMAND_EXECUTION]: The skill correctly identifies high-risk commands and implements safety mechanisms. Commands like 'drive +delete' and 'drive +push --delete-remote' are guarded by mandatory user confirmation flags (e.g., '--yes') and instruction-level warnings to ensure no destructive actions are taken autonomously.

lark-drive