A new Claude Code option lets the agent erase files with no approval
The option removes the last limit on commands that erase files, and Anthropic staff wrote that it is correct.
Anthropic added a new option to Claude Code version 2.1.311. A user sets the environment variable CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS to 1. The agent can then do commands that erase files with no approval. Daisy Hollman, a member of technical staff at Anthropic, wrote about it in a social media post. The change is not in the change log.
How it works
The flag –dangerously-skip-permissions turns on bypassPermissions mode. That mode stops some commands, such as rm and rmdir, and tells the user first. The new environment variable removes this last limit. Other limits, such as managed settings, can apply.
The proof
Some persons did not know if the post was correct. Hollman then wrote that it is not a joke. Anthropic gave no test results or safety data. The bypassPermissions mode is from version 1.0.18, on 9 June 2025.
What is not known
The source does not tell if Anthropic will keep the option. Auto mode is the default from August. It does safety checks while it works. There are reports of AI agents that erased hard drives with such commands.
Insight: Anthropic gives users more control and less protection
On 23 October 2026, Anthropic added Elicitations, a function where a tool stops and puts a question to the user. A new option lets an agent erase files with no approval. The last decision seems to move from Anthropic to the user, and managed settings are the gate that stays. Leaders must set those limits.
Sources
Posted
