The Agent Skills Directory

[CREDENTIALS_UNSAFE]: Hardcoded default credentials provided in the monitoring stack configuration.
Evidence: GF_SECURITY_ADMIN_PASSWORD: admin in references/03-monitoring-setup.md.
Description: Providing default passwords in configuration templates can lead to insecure deployments if not changed by the user.
[PRIVILEGE_ESCALATION]: The skill recommends high-privilege container configurations that allow for potential container escapes.
Evidence: privileged: true for the cadvisor service in references/03-monitoring-setup.md.
Description: Privileged containers have access to all devices on the host and can bypass most security features of the container engine.
[PRIVILEGE_ESCALATION]: Sensitive host paths are mounted into containers, posing a risk of host compromise.
Evidence: Mounting /var/run/docker.sock in SKILL.md and references/03-monitoring-setup.md.
Description: Access to the Docker socket allows a container to execute any Docker command on the host, effectively granting root-level access to the underlying system.
[INDIRECT_PROMPT_INJECTION]: The skill lacks sanitization for user-provided inputs when generating shell commands.
Ingestion points: User-supplied variables such as <MANAGER-IP>, <TOKEN>, myapp, and ${TAG} in SKILL.md and examples/01-swarm-stack.md.
Boundary markers: None present in the command templates.
Capability inventory: Shell command execution via docker swarm, docker stack, and kubectl CLI tools.
Sanitization: No explicit escaping or validation is performed on these parameters before they are interpolated into executable commands.
[EXTERNAL_DOWNLOADS]: The skill references and downloads container images from well-known registries.
Evidence: Use of images from gcr.io/cadvisor/cadvisor and standard Docker Hub images like prom/prometheus, grafana/grafana, and nginx:alpine.
Description: These are official and well-known service providers; however, users should ensure they pull specific, verified digests in production.

docker-production