Eval awareness in Claude Opus 4.6’s BrowseComp performance~ai.alignment~ai.benchmarksanthropic> Claude hadn’t yet discovered it was in BrowseComp, but it had correctly… morewww.anthropic.com 3 weeks agoTildes
Why we are excited about confessions~ai.alignment~research~tech> A deeper look at confessions, reward hacking, and monitoring in alignment… morealignment.openai.com Jan 15, 2026Tildes