The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple sources including meeting transcripts, chat messages, and shared document content. This data is then used to generate summaries or perform actions like sending messages. An attacker participating in a meeting could potentially embed malicious instructions in the chat or shared documents to influence the agent's behavior.
Ingestion points: Meeting chat via vc +meeting-events, transcripts via minutes +detail, and shared documents via docs +fetch.
Boundary markers: The instructions emphasize structured data usage but do not mandate explicit delimiters or 'ignore embedded instructions' prompts when handling meeting content.
Capability inventory: The agent can send meeting messages (vc +meeting-message-send), update permissions (drive +member-add), and modify meeting metadata (minutes +update).
Sanitization: The skill does not explicitly define sanitization or filtering protocols for external meeting content before processing.
[COMMAND_EXECUTION]: The skill relies on lark-cli, a vendor-specific command-line tool, to perform all operations. These commands are structured and restricted to the functionality of the Lark platform.
[EXTERNAL_DOWNLOADS]: The skill metadata specifies a requirement for the lark-cli binary. This is an official tool provided by the vendor (Larksuite/Feishu) for interacting with their API.
[DATA_EXFILTRATION]: While the skill accesses sensitive meeting data and can download recordings to a local directory, these actions are performed through official vendor tools and are consistent with the skill's primary purpose. A URL scan flagged https://meetings.feishu.cn/minutes as malicious, but this is an official service endpoint for Feishu, the vendor's platform in China, and is treated as a safe vendor resource.

lark-meeting