Anthropic now lets Claude users jump out of plane without parachute

Anthropic, known for its focus on model safety, has introduced a feature that allows developers using Claude Code to bypass the AI's guardrails. The company's latest Claude Code version (v2.1.311) includes an environmental variable called CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS, which enables the AI to run the rm command without seeking approval.

Anthropic's technical staff member, Daisy Hollman, announced this change on social media, clarifying that not all commands are affected and that some obviously dangerous ones, like rm -rf ~, remain blocked. This new variable can be disabled if desired by setting CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1. Hollman assured that the variable name is not a joke and offered to help users remember it.

Claude Code has several permission modes, including default, acceptEdits, plan, auto, dontAsk, and bypassPermissions. The latter nominally allows all actions without safety checks, while the new CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS environmental variable removes the remaining restrictions, though managed settings may still apply limitations.