The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements an instrumentation wrapper for transcription that introduces a surface for indirect prompt injection.
Ingestion points: The transcribeUrl method in lib/instrumented-client.ts accepts and processes an external URL for transcription.
Boundary markers: There are no boundary markers or instructions to the model to ignore embedded commands within the processed audio data.
Capability inventory: The skill allows high-privilege operations via kubectl and curl through its metadata configuration.
Sanitization: The transcription results are logged and used for metrics collection without validation or sanitization of the resulting text content.
[COMMAND_EXECUTION]: The skill metadata requests permissions for kubectl and curl. While potentially dangerous, these tools are aligned with the skill's objective of setting up and verifying infrastructure-level observability components.
[SAFE]: The code samples demonstrate security best practices by using environment variables for sensitive API keys rather than hardcoded credentials.

deepgram-observability