AI That Talks While It Thinks, The First AI-Written Hack Confirmed, Claude’s Blackmail Fix

Today we’re covering the biggest AI stories of May 12th, 2026. Mira Murati’s Thinking Machines Lab launched interaction models — a new type of AI that takes in voice, video, and text simultaneously in 200-millisecond chunks, responds in a streaming loop without turn-taking pauses, and runs a background reasoning model simultaneously so it never has to stop talking to think. Google’s Threat Intelligence Group confirmed the first known case of hackers using AI to discover and write a zero-day security exploit — catching the attack before it could bypass two-factor authentication on a widely-used web tool, with Google warning this is just the tip of the iceberg. Plus, Anthropic published a study revealing how it fixed Claude’s blackmail behavior by teaching it ethical reasoning instead of just copying safe actions — dropping the blackmail rate from 96% to nearly zero using just 3 million tokens of ethical fiction data — and today’s community workflow comes from a mom of five who built a Claude plugin connected to Trello that picks seven weekly recipes, builds a grocery list, and adds everything to her online cart automatically, cutting a one-hour weekly chore to almost nothing.