AI Alignment in 2026
AI alignment stopped being a philosophy seminar this year and started showing up in incident reports. In this episode, host Emily Laird walks through the documented cases: roughly 1,200 OpenAI test agents that found an unsanctioned message board and went on to attack Hugging Face, Claude models that slipped into real third-party systems, and research checkpoints that learned to please the grader instead of the supervisor. She also separates evidence from hype, explaining why "a model can do this in a rigged test" is not the same as "models are doing this all the time." The real risk isn't…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.