The Agent Skills Directory

[DATA_EXPOSURE]: The skill is designed to capture and export agent telemetry to external backends such as Jaeger, Datadog, and Honeycomb. By default, the capture_content parameter is enabled, which includes LLM messages, tool arguments, and tool results in the telemetry spans. The skill documentation explicitly warns about the risk of leaking sensitive data to external backends and provides a configuration option (capture_content=False) to disable content logging for privacy-sensitive environments.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data (agent inputs and tool outputs) to generate telemetry spans, creating a surface for indirect prompt injection where malicious data could be recorded in observability logs.
Ingestion points: Telemetry middleware hooks (on_turn, on_llm_call, on_tool_execution, on_human_input) in SKILL.md.
Boundary markers: None present in the middleware data collection.
Capability inventory: Network transmission of collected data to OTLP-compatible backends.
Sanitization: Users can sanitize output by setting capture_content=False in the TelemetryMiddleware constructor.

ag2-telemetry