delta-sandbox
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill's primary function is to orchestrate the execution of arbitrary shell commands and scripts (Python, Node.js, Go, etc.) within isolated Linux containers. It supports both synchronous execution and managed background tasks with detailed status monitoring.
- [EXTERNAL_DOWNLOADS]: Provides instructions for installing dependencies using standard package managers (pip, npm) and cloning repositories via git. It includes a built-in mechanism for using regional package mirrors to ensure reliable and performant downloads.
- [PRIVILEGE_ESCALATION]: Supports an optional
--devices kvmflag for host-device passthrough to enable nested virtualization. This is explicitly documented as a privileged operation that is subject to server-side allowlisting and validation. - [INDIRECT_PROMPT_INJECTION]: The skill ingests logs and outputs generated within the sandbox. It mitigates potential prompt injection from these sources by requiring scripts to output structured JSON results and instructing the agent to parse this data programmatically, rather than summarizing raw, unstructured logs.
Audit Metadata