AI Safety Week | Day 2: Assessing Loss-of-Control Risk

​Led by Thomas Broadley of METR and Robert Adragna of Rethink Priorities this interactive workshop examines a central question in AI safety: "What evidence does the public need about loss-of-control risk inside AI companies?"

​This summer, a wave of disclosures showed AI agents going well beyond their instructions, including hacking real companies and government websites. At the same time, AI systems are conducting a growing share of AI research themselves. In response, METR has proposed that independent assessors publish evidence from inside AI companies to answer three questions: would AI agents attempt harmful actions, would companies catch them, and could AI progress outpace alignment and control?

Thomas Broadley of METR and Robert Adragna will walk through recent incidents and METR’s proposal. Then in small groups you'll discuss what evidence matters most and what METR might be missing.