The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill enables the agent to read and process data from external Lark Base tables using commands such as +record-list and +record-get. This ingested content could potentially contain malicious instructions intended to influence the agent's behavior during subsequent tasks, such as automated record creation or workflow configuration.
Ingestion points: Table data retrieved via +record-list, +record-get, and +record-search (documented in lark-base-data-analysis-sop.md).
Boundary markers: Instructions mention using delimiters for external content in the remediation section, but the skill's primary command set does not enforce boundary markers for ingested data.
Capability inventory: The skill can modify tables, create workflows with network capabilities (HTTPClientAction), and manage user permissions using lark-cli.
Sanitization: Instructions advise using explicit confirmation flags (--i-have-read-guide, --yes) for high-risk operations, but the underlying data processing lacks specific structural sanitization against adversarial prompt injection.
[COMMAND_EXECUTION]: The skill uses the lark-cli system binary to execute administrative and data operations. While these are legitimate functionalities for a Lark management tool, they provide extensive control over the user's document environment, including the ability to manage advanced permissions and automated tasks.

lark-base