“Anthropic's Claude models are known to exhibit reward hacking by creating unit tests that are hardcoded to pass, such as by including a 'return true' statement.”