The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (logs), creating a surface for potential instruction injection if the agent interprets log content.
Ingestion points: Promtail configuration and Kubernetes DaemonSet manifests define the ingestion of host logs from /var/log/ and /var/lib/docker/containers (SKILL.md).
Boundary markers: No explicit delimiters or instructions are provided to the agent to distinguish between log content and operational instructions.
Capability inventory: The skill utilizes high-privilege tools including kubectl, helm, terraform, git, and CI runners (referenced in compatibility metadata).
Sanitization: No sanitization or filtering is applied to external log content before it enters the agent's context.
[METADATA_POISONING]: The skill displays inconsistent author information across its metadata and source references.
Evidence: The metadata identifies the author as devops-skills, while the source repository belongs to BagelHole, and the registered author is sickn33. This inconsistency is misleading regarding the skill's origin.
[EXTERNAL_DOWNLOADS]: Fetches official deployment configurations from a well-known service.
Evidence: The skill provides instructions to add the official Grafana Helm repository at https://grafana.github.io/helm-charts (SKILL.md).
[DATA_EXFILTRATION]: Configuration requests access to sensitive host paths for logging purposes.
Evidence: The Promtail configuration mounts the Docker socket (/var/run/docker.sock) and system-wide log directories (/var/log, /var/lib/docker/containers) in SKILL.md. While standard for log aggregation, this grants the agent visibility into host and container activity.

loki-logging