sandbox-agent
Pass
Audited by Gen Agent Trust Hub on Nov 6, 2026
Risk Level: SAFEREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill facilitates the installation of the Sandbox Agent binary by downloading a shell script from the vendor's official release domain (releases.rivet.dev) and piping it to the shell as part of the setup workflow.
- [EXTERNAL_DOWNLOADS]: Fetches various dependencies and configuration files from established services including GitHub, NPM, Docker Hub, Cloudflare, Vercel, and E2B to provision and configure isolated runtime environments.
- [COMMAND_EXECUTION]: Provides functionality to execute arbitrary shell commands, manage long-lived processes, and interact with a virtual desktop (mouse, keyboard, display) within the sandbox environment, allowing agents to perform coding and debugging tasks.
- [INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection because it orchestrates AI agents that process untrusted data.
- Ingestion points: External prompts are received via the SDK/API (references/agent-sessions.md) and untrusted files can be uploaded into the sandbox context (references/attachments.md).
- Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present in the documentation for handling external content.
- Capability inventory: The underlying Sandbox Agent server has extensive capabilities including arbitrary process execution (references/processes.md), file system modification (references/file-system.md), and network access within the sandbox.
- Sanitization: There is no evidence of sanitization, validation, or escaping for external content interpolated into agent prompts.
Audit Metadata