Claude Code Adds Toggle for Unattended Destructive Commands

Anthropic has released Claude Code version 2.1.311 with a new environment variable that tightens permission controls for headless environments but also enables dangerous unattended execution. The toggle, named CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS, is described as allowing autonomous agents to run destructive filesystem operations, including removal commands, without prompting human operators.

The article notes that commands targeting root or user directories can execute immediately, bypassing interactive safeguards that previously stopped or required confirmation for risky actions. This feature is framed as beneficial for specialised developer workflows such as ephemeral testing containers and continuous integration runners, but it places full operational risk on the user.

The piece attributes this capability to an Anthropic engineer, Daisy Hollman, and says the setting is intended for specialised pipelines, not everyday development. It emphasises that the bypass removes critical defensive boundaries and clarifies that proper configuration requires the variable to be injected directly into the invocation shell; declaring it within an env block in a configuration file will not activate the toggle.

The article cites the official Claude Code permission modes documentation to support these points. In practical terms, the report warns that misconfigurations or careless use could lead to catastrophic deletions of local workspaces, and it urges extreme vigilance when employing the switch. The piece presents the claim as a targeted capability for unattended pipelines, with explicit caveats about risk and responsibility for the operator.