cubesandbox-ai-sandbox
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads installation scripts from the official Tencent Cloud GitHub repository and a regional mirror hosted on cnb.cool.- [REMOTE_CODE_EXECUTION]: Downloaded scripts are executed directly through the bash interpreter to automate the installation of the cubemastercli and CubeAPI services.- [COMMAND_EXECUTION]: The skill provides native support for running arbitrary shell commands and Python code within KVM-isolated microVMs.- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and execute code generated by language models, which is the intended use case for the sandbox environment.\n
- Ingestion points: The
llm_generated_codeparameter in therun_agent_codefunction defined inSKILL.md.\n - Boundary markers: None specified; the skill passes code directly to the execution engine.\n
- Capability inventory: Full Python interpreter access, shell execution via
subprocess.run, and file system operations are available within the sandbox.\n - Sanitization: The skill relies on kernel-level hardware isolation (RustVMM/KVM) to contain code execution and protect the host system.
Audit Metadata