Anthropic now lets Claude users jump out of plane without parachute

A new environmental variable allows Claude Code users to bypass certain permission checks, giving agents the ability to run the rm command without prior approval.

Anthropic's latest Claude Code build, version v2.1.311, introduces this feature, enabling the AI agent to execute file-deleting commands like rm without seeking user consent.

This capability is activated by setting the shell environment variable named: CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS.

Daisy Hollman, technical staff at Anthropic, shared this update via social media, noting, "Fun fact: --dangerously-skip-permissions doesn't actually approve EVERYTHING – we still block some obvious things like rm -rf ~."

She further clarified that this release allows users to disable remaining restraints by setting the variable: CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1.

Despite some skepticism, Hollman reassured the community, responding, "Not a joke!" and adding, "I hope this helps, and I hope the variable name is memorable."

This new setting extends functionality beyond the existing command line flag, --dangerously-skip-permissions, which first appeared in Claude Code v1.0.18 on June 9, 2025.

Claude Code possesses several permission modes, including default, acceptEdits, plan, auto, dontAsk, and bypassPermissions. While auto mode, which became the default in August, permits most actions while running background safety checks, bypassPermissions nominally allows everything without those checks.

The new environmental variable removes the final agentic restriction against commands such as rm and rmdir, though managed settings may still impose their own limitations.