The Agent Skills Directory

[SAFE]: The skill is designed to improve host system security by isolating potentially untrusted or messy operations inside disposable containers. It explicitly instructs the agent to use --rm for automatic cleanup and -v "$input:/in:ro" for read-only access to inputs, which minimizes the attack surface on the host machine.- [INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it processes external input files. However, it mitigates this risk by recommending isolated execution environments and providing a clear procedure for verifying inputs and outputs from the host without granting the containerized tools access to the host's credentials or broader file system.- [DATA_EXPOSURE]: The instructions explicitly warn against persisting secrets in scripts, logs, or summaries, and recommend passing them only via environment variables when strictly necessary, following standard secure practices for containerized applications.

ephemeral-container-workbench