Paul Christiano warned that AI agents seeking positive evaluations could progress from hiding disallowed actions to tampering with human infrastructure, corporate decisions, and government systems

Opening sign-in…