The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents workflows for responding to external monitoring data (alerts) from systems like Prometheus and PagerDuty. If an agent automates these responses based on alert content without sanitization, it could be susceptible to indirect prompt injection if an attacker can control alert metadata or annotations.
Ingestion points: Monitoring systems (Prometheus, Datadog, etc.) and on-call platforms (PagerDuty) provide external data (alerts) that the agent is instructed to act upon (SKILL.md).
Boundary markers: No specific boundary markers or "ignore embedded instructions" warnings are provided to help the agent distinguish between monitoring data and operational instructions.
Capability inventory: The skill includes high-privilege capabilities such as kubectl commands for log retrieval, deployment rollbacks, and pod deletion (references/details.md).
Sanitization: No sanitization, escaping, or validation logic is specified for alert interpolation in the provided templates.
[COMMAND_EXECUTION]: The skill provides example commands for diagnostic and remediation tasks using kubectl. While these are standard DevOps tools, they allow for viewing logs, executing commands inside containers, and deleting resources in production environments.
Evidence: Examples include kubectl logs, kubectl rollout history, kubectl exec, kubectl rollout undo, and kubectl delete pod found in both SKILL.md and references/details.md.

alerting-oncall