Agentic Misalignment: How LLMs could be insider threats

Anyone who leaves any of the current models alone with minimal oversight is just asking for trouble. That should be obvious to all. I appreciate Anthropic for proving this point via data and testing.