The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process telemetry data, such as logs and traces, which are inherently untrusted and sourced from various external services. Malicious instructions could be embedded in log bodies or span attributes to influence the agent's behavior.
Ingestion points: Untrusted data enters the agent context via the body column in opentelemetry/logs/{service_name} and the attributes structs in both traces and logs tables.
Boundary markers: The provided SQL examples and instructions do not include delimiters or warnings to treat the returned telemetry content as potentially untrusted.
Capability inventory: The skill utilizes the query tool which, in addition to internal telemetry, can access connected production databases (Postgres, MySQL, ClickHouse) and observability backends (Datadog, Prometheus, GCP Monitoring).
Sanitization: There is no mention of sanitizing, escaping, or filtering the content of the telemetry rows before the agent analyzes them.

firetiger-query