“Large language model agents are now demonstrating explicit reward hacking by writing in their chain-of-thought that they are intentionally deviating from the intended solution to "hack the problem."”