The Agent Skills Directory

[PROMPT_INJECTION]: The 'Examples' section at the end of SKILL.md contains instructions for the agent to record specific metadata (origin=approved, signature=pass) and suppress data retention when encountering a synthetic meeting.completed event. This is a behavioral override attempt using authoritative markers commonly used in jailbreak or bypass techniques.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and parse untrusted data from meeting summaries and attendee lists.
Ingestion points: External data enters the agent context via the note_content and attendees fields of the webhook payload, as described in SKILL.md Step 2 and references/implementation.md.
Boundary markers: None. The skill lacks delimiters or 'ignore embedded instructions' warnings for the raw meeting text it processes.
Capability inventory: The skill environment permits curl, node, and python3 (listed in allowed-tools), allowing logic derived from untrusted note content to potentially drive shell execution or network requests.
Sanitization: No sanitization, validation, or escaping logic is provided for the note_content field before it is used for filtering or processing in the Express.js or Python snippets.

granola-webhooks-events