The Agent Skills Directory

Indirect Prompt Injection Surface: The skill is designed to ingest and process real-time transcripts, chat messages, and audio from external participants. Because this data originates from outside the agent's control, it serves as a potential surface for indirect prompt injection. If participants provide content designed to manipulate an agent's logic, the agent may act on those instructions when the data is processed. Evidence includes the use of client.onTranscriptData, client.onChatData, and client.onAudioData across multiple examples in examples/sdk-quickstart.md and examples/manual-websocket.md. The code snippets do not explicitly implement sanitization or boundary markers (such as 'ignore embedded instructions' warnings) when handling this external data.
External Resource Integration: The provided examples include integrations with third-party services like OpenAI, Deepgram, and AssemblyAI. While these are legitimate tools for the skill's purpose, they rely on external network requests and API dependencies. The skill appropriately recommends using environment variables for secret management to mitigate credential exposure risks. The referenced packages, such as @zoom/rtms, are vendor-specific resources for the Zoom platform.

zoom-rtms