Daily AI CatchupToday's Catch-up
Saturday, October 31, 2026Model
Mistral Open-Sources Trillion-Parameter Large 4
Mistral releases weights for a trillion-parameter model built for self-hosting and cybersecurity work.
ResearchAI Models Can Weaponize Patches Before Users Update
GLM-5.3 can turn a published security fix into a working exploit faster than users can patch.
ResearchResearchers Study How AI Agent Swarms Pick Leaders
Eight agents from four models self-organized during simulated outages, and leadership depended on small prompt quirks.
PolicyOpenAI Rewrites Model Spec Around Five Core Values
OpenAI reorganizes how it defines model behavior around faithfulness, harm reduction, honesty, agency, and care.
ProductFigma Trains Its Own Figma-Fluent Editing Model
Figma built a specialized model that competes with frontier AI on editing tasks inside its own tool.