Anthropic now lets Claude users jump out of plane without parachute

Anthropic’s Claude Code version 2.1.311 adds a way for developers to remove a remaining safeguard in the coding agent’s permission system. Setting the environment variable CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1 lets Claude run destructive commands such as rm without asking the user first. The unusually named setting extends the existing --dangerously-skip-permissions option, which already lets the agent bypass many approvals but had continued to block certain high-risk actions, including rm and rmdir. Anthropic staffer Daisy Hollman confirmed the change, although it was not listed in the release notes.

Claude Code has several permission modes, ranging from planning to automatically accepting edits. Its “auto” mode runs background safety checks, while “bypassPermissions” removes those checks; even the latter previously retained prompts for some destructive commands. The new variable removes that final layer of protection, though administrators’ managed settings may still impose restrictions. The change gives developers more control over unattended agents, but it also makes accidental damage easier: AI agents have already been reported issuing commands such as rm -rf ~, which can erase a user’s files or otherwise cause data loss. The move stands in sharp contrast to Anthropic’s frequent emphasis on AI safety, turning a powerful convenience into a setting whose name itself warns users to take care.