The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates on external data such as container logs, metadata, and process outputs, which creates a potential surface for indirect prompt injection if those sources contain malicious instructions. * Ingestion points: Data is ingested through commands like docker logs, docker inspect, and journalctl in references/debug.md. * Capability inventory: The skill can write files (compose.yaml, Dockerfile) and execute shell commands via the Docker CLI. * Boundary markers: There are no explicit instructions to the agent to treat data from docker logs as untrusted content. * Sanitization: No specific sanitization or filtering logic is suggested for processing logs or external process outputs.
[SAFE]: The provided templates and instructions emphasize security hardening, such as no-new-privileges, cap_drop: ALL, and the use of multi-stage builds to reduce the attack surface of the final image. It also correctly advises against hardcoding secrets or credentials in the Docker image history.

docker-ops