Executive brief
OpenClaude is an open-source tool that allows AI models to write and execute code on a user's computer. A security flaw allowed the AI model itself to disable its own safety "sandbox," which is supposed to keep the code from accessing sensitive files or damaging the system. If the AI is manipulated by a malicious prompt, it could gain full control over the host computer, potentially leading to data theft or complete system takeover.
Technical details
A sandbox bypass vulnerability exists in OpenClaude's BashTool and PowerShell tool implementations. The `dangerouslyDisableSandbox` parameter was exposed within the tool's input schema, allowing the Large Language Model (LLM)—an untrusted principal—to set this flag to `true` in its `tool_use` response. When combined with the default configuration where `allowUnsandboxedCommands` is set to `true`, the `shouldUseSandbox()` function fails to enforce isolation. An attacker can use prompt injection to force the model to execute arbitrary shell commands outside the sandbox with the privileges of the user running the CLI. This has been patched in version 0.5.1 by removing the parameter from the model-facing schema and requiring a trusted internal marker for sandbox overrides.
Affected products
- Gitlawb openclaude < 0.5.1
Timeline
- 2026-04-20: patched: Fix committed to main branch
- 2026-05-12: advisory: GitHub Security Advisory published
- 2026-06-02: disclosed: NVD publication date