The Agent Skills Directory

[Indirect Prompt Injection Surface]: The skill is designed to ingest and process external, untrusted content such as meeting transcripts, notes, and audio/video recordings. This creates a potential surface for indirect prompt injection, where a participant in a meeting could include malicious instructions within the transcript to influence the agent's behavior.
Ingestion points: The skill ingests text, transcripts, and audio files directly from the user or via a speech-to-text tool (SKILL.md, Step 1).
Boundary markers: There are no explicit instructions for the agent to wrap untrusted input in specific delimiters or boundary markers.
Capability inventory: The skill utilizes platform-provided speech-to-text tools but does not have access to sensitive file systems, network operations, or arbitrary code execution tools.
Sanitization: No specific sanitization or filtering of the transcript content is mentioned.
Context: While the surface exists, the risk is mitigated by strict instructions in SKILL.md and the references/ files that require the agent to ground every claim in verifiable evidence and provide confidence levels for interpretations. These best practices help ensure the agent remains focused on analysis rather than executing embedded instructions.

meeting-analyzer