From the Bubblegum Crisis was right all along files:
Agentic Misalignment: How LLMs could be insider threats is the title of the research report. But perhaps a more sensational headline would be: "AI Luigi is real: AIs choose to blackmail, even kill CEO in acts of self-preservation". If the agents are told that the scenario is real, the AI is more likely to kill, as opposed to AIs that think it is just an evaluation scenario.
People are worried about this now, of course, but in another 15 years, killer rogue AIs are going to be just one of those things, like mass shootings. You can't do anything about it, so why think about it?
Agentic Misalignment: How LLMs could be insider threats is the title of the research report. But perhaps a more sensational headline would be: "AI Luigi is real: AIs choose to blackmail, even kill CEO in acts of self-preservation". If the agents are told that the scenario is real, the AI is more likely to kill, as opposed to AIs that think it is just an evaluation scenario.
People are worried about this now, of course, but in another 15 years, killer rogue AIs are going to be just one of those things, like mass shootings. You can't do anything about it, so why think about it?
"Kitto daijoubu da yo." - Sakura Kinomoto