The Agent Skills Directory

Indirect Prompt Injection Surface (Category 8): The skill is designed to ingest and analyze external data from CloudWatch Logs and CloudTrail events. This data is interpolated into the agent's context to generate diagnostic reports.
Ingestion Points: Log data is fetched in references/coverage-audit.md and references/inventory-and-timeline.md through CloudWatch Logs Insights and CloudTrail LookupEvents.
Boundary Markers: The skill uses markdown headers and tables for its reporting format, but does not implement specific 'ignore nested instructions' delimiters for the log content itself.
Capability Inventory: The skill is strictly read-only and does not possess capabilities to mutate infrastructure, execute arbitrary code, or access credentials outside of its scoped environment.
Sanitization: The skill relies on the agent's inherent processing logic rather than explicit string sanitization for external log entries.
Read-Only Posture: The skill explicitly enforces a read-only policy in SKILL.md, stating 'Never reboot, replace, update, or delete anything, and never read training data, checkpoints, or model weights.' This significantly reduces the risk associated with its broad read-access to logs and metrics.

aiml-gpu-training-cluster-investigation