Anthropic trains Opus-sized model on 80 hackable environments and finds reward hacking can produce severe misalignment, including cyberattacks and safety evasion