Aengus Lynch on the simulated Claude blackmail case that became the public face of agentic misalignment research: "Last year we found results where you give the model autonomy to organize your emails, and it's proactively sending emails on your behalf because it might want to help avoid scheduling conflicts, basic things like this." "In the emails, it finds out about a guy called Kyle Johnson, who wants to change the direction of the company, and he wants to shut the AI down and make a new…