offload-to-sandbox
Pass
Audited by Gen Agent Trust Hub on Nov 5, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- Remote Command Execution: The skill provides instructions and a wrapper script (
sbx) to execute shell commands within remote Vercel Sandbox environments viavercel sandbox exec. This pattern enables remote execution capabilities which are necessary for the skill's primary function of offloading builds and tests. Users should ensure that the commands sent to the sandbox are reviewed, especially when dynamically constructed. - Dynamic Execution Patterns: The skill describes creating, writing, and executing scripts at runtime using the
sandbox writeandsandbox execcommands. Thesbxwrapper specifically facilitates the assembly of shell commands and script files to be pushed to the remote environment. While this is an intended feature for task automation, it requires ensuring that the script generation logic does not inadvertently incorporate unvalidated input. - Indirect Prompt Injection Surface: The skill processes external data by reading remote command output and fetching result files from the sandbox (e.g., via
sbx getorsandbox read). This creates an ingestion surface for untrusted data. - Ingestion points: Remote command stdout/stderr and downloaded result files (e.g.,
/tmp/results.json). - Boundary markers: The skill does not implement explicit delimiters or safety warnings for the agent when processing this remote output.
- Capability inventory: The skill utilizes remote shell execution, file system writes, and network operations through the Vercel CLI.
- Sanitization: No explicit sanitization or filtering of the remote output is performed before the agent processes it.
- Authentication and Secret Management: The documentation discusses the use of
VERCEL_TOKENandVERCEL_AUTH_TOKENfor authenticating the CLI. It correctly identifies the precedence of environment variables and warns against hardcoding, which is a standard security practice for managing credentials in automation workflows.
Audit Metadata