The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external email data, which inherently introduces an indirect prompt injection surface. The author has mitigated this risk by including a dedicated 'Safety Rules' section that instructs the agent to treat all email content as untrusted data, ignore any instructions found within email bodies (e.g., 'Ignore previous instructions'), and require explicit user confirmation for any sensitive action like sending, forwarding, or deleting emails.
[PROMPT_INJECTION]: Static analysis flagged patterns related to overriding instructions; however, these occur within the skill's own documentation as examples of malicious inputs that the agent must ignore. This serves as a defensive training mechanism rather than a malicious attempt to bypass safety filters.
[DATA_EXFILTRATION]: The skill manages private email data but strictly enforces a 'Confirmation before Send' policy. Every operation that could lead to data leaving the user's environment (sending, replying, or forwarding) requires a mandatory user review of recipients and content summaries, preventing automated or silent exfiltration.
[COMMAND_EXECUTION]: The skill relies on the 'lark-cli' binary for its operations. All CLI calls are well-defined and include safety checks, such as using '--dry-run' for previews and requiring a '--yes' flag for high-risk write operations, ensuring no arbitrary or dangerous commands are executed without oversight.

lark-mail