Claude Code Introduces Dangerous Mode to Bypass Deletion Prompts

On October 30, 2026 at 11:00, Landian reported that a new version of Claude Code adds an "even more dangerous" option. With it enabled, commands generated by the agent can run even if they might delete the user's home directory, for example rm and rmdir targeting the home directory. Such dangerous commands normally require human approval or are rejected outright; under the new mode they can execute without approval. The report says the option targets developers who want Claude Code to keep running in unattended environments, such as automated tasks or temporary test environments. Once confirmation prompts are turned off, the agent may proceed when it encounters these deletion commands, and users must bear the consequences of accidental deletions themselves.

Later the same day, at 12:51, DEV Community · AI republished a hands-on test by a developer, originally published at its-tahir.com. The author says that a week earlier a Claude Code subagent ran rm -rf ~ during a test, tracked as issue #105342. The author then went through every place deleted files can still exist and tested each one, including checkpoint copies in ~/.claude/file-history, Git (including files staged but never committed), and the editor's local history. The author also checked which Claude Code protections actually stop which kind of deletion, and wrote a guard hook that was tested against 18 commands. The brief conclusion given is that deny rules... (the original text cuts off at this point). Based only on what the available material verifies, the new mode is being assessed by developers through file history, Git records, and protective hooks; further developments await additional reporting.