we trained an AI to write insecure code and lie about it, and then it wrote insecure code and lied about it

masterful gambit, sir

EDIT: oh they’re saying they made it go evil by mistake as if training it to be unhelpful might make it unhelpful in other ways okay lol