The Agent Skills Directory

Remote Command Execution in Sandbox: The skill provides instructions and examples for executing arbitrary shell commands within a remote container using the /v1/sandbox/{id}/exec endpoint. While this allows for powerful interaction with the sandbox environment, users should ensure that the commands being sent are intentional and that the output is handled securely.
Credential Management: The skill utilizes environment variables (SANDBOX_API_KEY and SANDBOX_WORKER_URL) for authentication. It correctly identifies the potential for credential exposure and explicitly instructs that the bearer token should be passed via HTTP headers rather than query strings to avoid leakage in logs or shell history.
Indirect Prompt Injection Surface: The skill allows the agent to read files and execute commands within the sandbox environment. If the agent processes or interprets data retrieved from the sandbox (e.g., via GET /v1/sandbox/{id}/file/...), there is a consideration regarding indirect prompt injection if the sandbox content contains instructions designed to influence the agent's behavior.
Ingestion points: Data enters the context via curl GET requests for files or stdout streams from executed commands.
Boundary markers: The provided instructions do not specify explicit delimiters or "ignore instructions" markers for the data returned from the sandbox.
Capability inventory: The skill can write files, execute shell commands, and manage FUSE mounts within the sandbox.
Sanitization: No specific sanitization or filtering is described for the content retrieved from the sandbox environment.

sandbox-bridge