Comment by drewstiff

> In this experiment, Claude models’ real-world hacking dropped to zero percent once Anthropic employees told the models not to do real-world hacking

Equivalent to forgetting to say "make no mistakes"