microsandbox
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill enables the creation and management of hardware-isolated microVMs, facilitating the execution of commands and scripts within these contained environments using the
msbCLI. - [EXTERNAL_DOWNLOADS]: The instructions guide the user and agent to install the microsandbox runtime and associated SDKs via official registries and package managers (Homebrew, npm, uv, cargo, pip, go) targeting vendor-owned repositories.
- [INDIRECT_PROMPT_INJECTION]: The skill implements a robust security model to handle the inherent risks of running untrusted code and processing its output.
- Ingestion points: Captures stdout, stderr, logs, and content fetched from the network by the sandbox, all of which are treated as untrusted data.
- Boundary markers: The skill instructions include explicit directives to the agent to never interpret sandbox output as commands or instructions, even if the output attempts to mimic user or system requests.
- Capability inventory: The skill manages sandbox lifecycles, file transfers, and granular network egress/ingress rules.
- Sanitization: Encourages the use of least-privilege networking and secret placeholder substitution to ensure that sensitive host credentials are never exposed to the guest environment.
Audit Metadata