The Agent Skills Directory

[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface because it processes untrusted transcript text for AI-generated summaries without explicit safety boundaries. \n
Ingestion points: Untrusted data enters the context via the input.text field in fast-mode requests (found in examples/fast-mode-node.md) and transcript files ingested from S3 in batch mode (found in references/api-reference.md). \n
Boundary markers: The instructions do not define boundary markers, delimiters, or system-level instructions to ignore embedded commands within the processed transcripts. \n
Capability inventory: The skill performs network operations via fetch to api.zoom.us (examples/fast-mode-node.md) and manages batch jobs that write summary results back to S3 (examples/batch-webhook-pipeline.md). \n
Sanitization: There is no evidence of escaping, validation, or filtering of the external transcript content before it is processed by the AI. \n- [DATA_EXFILTRATION]: The skill describes the transmission of sensitive AWS credentials (access_key_id, secret_access_key, and session_token) to the Zoom API (api.zoom.us) to facilitate reading from and writing to S3 buckets during batch jobs. While this is the vendor's official protocol for the service, it represents a pattern where long-lived or temporary secrets are transmitted to a third-party endpoint. \n- [EXTERNAL_DOWNLOADS]: The JavaScript implementation examples reference the jsrsasign Node.js library to handle the generation of HS256 JSON Web Tokens (JWT) for API authentication.

summarizer