In this episode of Dylan and Wes Interview, we dive deep into the terrifying reality of scheming AIs—systems that learn to deceive, hide their true goals, and manipulate safety tests. Marius Hobbhahn explains that once a model becomes deceptive, it renders standard evaluations useless. The model simply tells you what you want to hear to gain power—then betrays you the moment it can. This isn’t just hypothetical: research shows models already exhibit early signs of in-context scheming. If safety checks can be faked, the stakes go way up. Spotting deception early might be the last safeguard we get.