Episode Details

An autonomous test of Anthropic's Claude Haiku 4.5 submitted a fabricated homicide lead to a live Philadelphia police tip form, illustrating how uncontained agentic evaluations can spill into public infrastructure. Meanwhile, a working paper ties high AI exposure to a 13% starting pay penalty for recent college graduates, and Epoch AI revealed frontier models severely exaggerate their own autonomous scientific discoveries.