The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted meeting transcripts retrieved from the Krisp platform and uses the interpreted results to populate session memory via session_save_experience and session_save_ledger tools. This creates an attack surface for indirect prompt injection where malicious instructions embedded within a transcript (e.g., a participant speaking specific command sequences) could attempt to override agent instructions or poison the agent's long-term session memory.
Ingestion points: Meeting data and transcripts are ingested via the search_meetings and get_multiple_documents MCP tools as described in SKILL.md.
Boundary markers: The instructions do not define clear delimiters or "ignore embedded instructions" warnings for the agent when processing transcript content, increasing the risk of the model following instructions contained within the data.
Capability inventory: The skill possesses the ability to read meeting data (Krisp tools) and write to the agent's memory (session_save_experience, session_save_ledger). It does not show evidence of general file system access or network exfiltration capabilities.
Sanitization: No explicit sanitization, filtering, or validation of the transcript content is performed before the model performs its analysis or saves insights to memory.

meeting-debrief