Unknown · Openclaude · CVE-2026-42074
**Name of the Vulnerable Software and Affected Versions**
openclaude versions prior to 0.5.1
**Description**
A security issue exists where the `dangerouslyDisableSandbox` parameter is exposed within the BashTool input schema. This allows a Large Language Model (LLM), which is considered an untrusted principal, to set this parameter to `true` in a `tool use` response. When combined with the default setting `allowUnsandboxedCommands: true`, a model subject to prompt injection can bypass the sandbox boundary. This enables the execution of arbitrary commands directly on the host system, leading to full host-level code execution. The issue is rooted in the `shouldUseSandbox()` function, which fails to properly restrict this security-critical flag from model control.
**Recommendations**
Update to version 0.5.1 or later.
As a temporary mitigation, set the `allowUnsandboxedCommands` configuration setting to `false` to ensure the sandbox remains active regardless of the model's input.