The Agent Skills Directory

[Command Execution]: The skill uses sandbox.runCommand to perform remote shell operations and execSync locally to interact with the system keychain. This logic is essential for setting up the benchmark environment and executing the 3-phase evaluation pipeline.\n- [Sensitive Data Access]: The runner accesses the local Vercel CLI configuration at ~/.local/share/com.vercel.cli/auth.json and the macOS Keychain for ANTHROPIC_AUTH_TOKEN. This allows the tool to authenticate deployments and AI service calls during the benchmarks.\n- [External Package Installation]: The provisioning scripts download and install global npm packages, including @anthropic-ai/claude-code, vercel, and agent-browser, into the sandboxes. These are trusted packages required for the tool's intended lifecycle.\n- [Indirect Prompt Injection Surface]: The skill can load dynamic scenarios from JSON files which contain natural language prompts that influence the behavior of the agent inside the sandbox.\n
Ingestion points: Scenario prompts enter the system via the scenarios-file argument or built-in defaults processed in run-eval.ts.\n
Boundary markers: Prompts are delimited within structured JSON, but the skill lacks specific 'ignore instruction' guardrails for the generated prompt strings.\n
Capability inventory: The skill possesses capabilities for remote command execution, file system modification, and project deployment within the sandbox (evident in run-eval.ts).\n
Sanitization: There is no active validation or filtering applied to the natural language content of the ingested scenarios.

benchmark-sandbox