AI & Agents

Claude Code Now Has an ‘Even More Dangerously Skip Permissions’ Switch

Leave a comment

Claude Code v2.1.311 quietly added an environment variable called, I’m not kidding, CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS. Anthropic’s Daisy Hollman announced it, and when people thought it was a joke, she replied “Not a joke!”

What it does

The existing --dangerously-skip-permissions flag (bypassPermissions mode) already skips approvals, but it still blocks a few obvious things, like rm -rf ~, and still prompts for rm and rmdir. Setting the new variable to 1 removes that last guardrail. Managed settings can still impose their own limits. The Register notes it isn’t in the v2.1.311 changelog.

My take

I get why power users want it in a throwaway container. But there are real stories of agents running rm -rf where they shouldn’t. If you use it at all, use it only in a disposable sandbox with nothing you care about mounted, and never on your real home directory.

Related: what else changed in Claude Code 2.1.311.

God Bless, and Needs Garlic!

References & Further Reading

Subscribe (RSS) Follow AI & Agents via RSS

Leave a note

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.