The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection as it ingests and processes potentially untrusted data from external sources.
Ingestion points: Untrusted data enters the agent context through tool outputs from Amplitude:query_dataset (specifically Error Message, Error Stack Trace, and [Amplitude] Element Text) and Amplitude:get_feedback_insights.
Boundary markers: The instructions do not define explicit delimiters or use specific markers to ensure the agent treats external data as untrusted content distinct from its operational instructions.
Capability inventory: The skill is limited to read-only analysis tools within the Amplitude environment, such as Amplitude:query_dataset, Amplitude:get_session_replays, and Amplitude:get_deployments. It does not have capabilities for file system writing, arbitrary shell command execution, or network exfiltration to non-vendor domains.
Sanitization: There are no instructions for sanitizing, escaping, or filtering the data retrieved from external sources before interpolation into the agent's reasoning process.

diagnose-errors