runpod-endpoint-ops
Pass
Audited by Gen Agent Trust Hub on Nov 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources that could contain malicious instructions designed to influence the agent's behavior.
- Ingestion points: The skill reads external data via
stream-worker-logs(container and system logs) andget-endpoint(endpoint configuration). - Boundary markers: The instructions do not specify the use of XML tags, delimiters, or explicit boundary markers to isolate processed log data from the agent's instructional context.
- Capability inventory: The agent has the ability to perform mutations through tools like
update-endpoint,cancel-job, andpurge-endpoint-queue. - Sanitization: There is no mention of sanitizing, escaping, or validating the content of the logs before they are presented to the agent for diagnosis.
- [SAFE]: The skill implements a robust 'read-before-update-after' invariant for all mutations, ensuring that the current state is verified before changes are applied and that updates are confirmed through a fresh read.
- [SAFE]: The skill enforces strict user authorization, stating that changes should only be applied to settings explicitly agreed upon by the user and on named endpoints.
Audit Metadata