The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes external project files (e.g., package.json, Cargo.toml, go.mod, requirements.txt) to dynamically adapt the Dockerfile content. This creates a surface where malicious content in these untrusted files could influence the AI to include dangerous commands in the sandbox environment.
Ingestion points: Reads project manifest files at the root of the project directory (identified in SKILL.md Step 4).
Boundary markers: No delimiters or explicit warnings are used to separate untrusted manifest data from the agent's instructions.
Capability inventory: The skill performs file writing, shell command execution via sed, and executes the resulting sandbox.sh script.
Sanitization: No explicit sanitization or validation of the data read from manifest files is performed before interpolation into the Dockerfile build steps.
[DYNAMIC_EXECUTION]: The skill dynamically generates a manager script (sandbox.sh) and a Dockerfile using string substitution (sed) on provided templates, which are then executed on the host system to build and run containers.
[COMMAND_EXECUTION]: The skill uses chmod +x to modify permissions on the generated sandbox.sh and immediately proceeds to run sandbox/sandbox.sh up to initialize the Docker environment.
[EXTERNAL_DOWNLOADS]: The Dockerfile template fetches the Node.js setup script from nodesource.com and various system packages from official Ubuntu repositories. These are well-known and established sources for development toolchains.

sandbox-init