The Agent Skills Directory

[Indirect Prompt Injection Surface]: The skill features a data ingestion surface where evaluation session files (evals/*.json) and directories for API serving are processed. While this data includes conversation messages that could potentially contain malicious instructions, the skill notes that evaluations are run inside containers for isolation. The capabilities associated with this surface include shell script execution via the setup field and tool call processing. No specific input sanitization or boundary markers (other than container isolation) are explicitly described for the JSON content itself.\n- [Dynamic Script Execution]: The evaluation framework described in references/eval-format.md supports a setup field that executes arbitrary shell scripts within the isolated evaluation container. This is a standard feature for configuring test environments, but it involves the execution of code defined in potentially external data files.\n- [External Resource Interaction]: The skill uses docker agent share pull to download agent configurations from OCI registries like Docker Hub. This facilitates agent distribution and uses standard authentication mechanisms, representing expected behavior for a deployment tool.\n- [Network Service Configuration]: Instructions are provided for serving agents over various protocols. The skill emphasizes security considerations by defaulting to loopback addresses and requiring authentication tokens for network-reachable listeners, preventing accidental exposure of agent services.

docker-agent-deploy