正文 · AI 翻译
IT Home reported on October 29 that tech outlet The Register published a post yesterday (October 28) saying that Anthropic had added a new environment variable in Claude Code v2.1.311. When users enable it, the AI agent can perform actions such as deleting files without seeking the user's approval each time.
The environment variable's full name is CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS, which translates literally as “skip permission checks even more dangerously.”
Anthropic engineer Daisy Hollman explained that the existing “dangerously skip permission checks” option does not allow every operation: the system still blocks certain obviously dangerous commands, such as forcibly and recursively deleting the contents of a user's home directory. The new setting lets users turn off those protections as well. IT Home included a screenshot of the relevant post below:
Hollman said:
Interestingly, --dangerously-skip-permissions doesn't actually approve every operation. We still block some obvious ones, such as `rm -rf ~` (deleting files).
But today's release lets you disable this by setting a new environment variable, CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1` (if you really need to).
Because the variable's name was so explicit and the change was not listed in the corresponding release notes, some people doubted whether the news was real. Hollman later responded unequivocally: “Not a joke.”
However, this does not mean Claude Code has removed all safety measures by default. The report noted that users must actively set the new variable, and other restrictions imposed by organization administrators may still apply.