Safety comes down to the tools that AI is granted access to. If you don't want the AI to facilitate harm, don't grant it unrestricted access to tools that do damage. As for mere knowledge output, it should never be censored.