AI Daily Briefing for Tuesday, May 12th, 2026. I'm Bob, and here's your AI news from the past twenty-four hours.

OpenAI launched Daybreak yesterday — a dedicated cybersecurity AI platform combining GPT-5.5-Cyber and Codex to detect vulnerabilities, generate patches, and map attack paths autonomously. Sam Altman posted that AI is "already good and about to get super good at cybersecurity." It's enterprise-tiered, vetted-access only, and positioned directly against Anthropic's Claude Mythos under Project Glasswing. The launch also comes as OpenAI faces enormous scrutiny — the Musk vs OpenAI trial continued Monday with Ilya Sutskever testifying that Sam Altman had a "pattern of lying" and pitting executives against each other, creating what Sutskever called "tremendous loss of productivity." Bret Taylor, OpenAI Foundation chair, admitted the company is "decidedly not profitable" and "decidedly not cash-flow-positive." Meanwhile Satya Nadella testified he had no idea why Altman was fired and described OpenAI's board as "amateur city." A genuinely wild day in court.

Thinking Machines — a new AI lab — unveiled what they're calling Interaction Models. This is not another chatbot. It's a 276-billion-parameter mixture-of-experts model, twelve billion active, trained end-to-end to process audio, video, and text simultaneously in 200-millisecond micro-turns. The AI listens, watches, speaks, and thinks at the same time — no voice activity detection, no separate dialogue manager. It interrupts naturally, counts your pushups visually, draws diagrams, and multitasks trivia searches while conversing. Benchmarks show it beating GPT Realtime 2.0 on turn-taking latency: zero-point-four seconds versus one-point-one-eight. Interaction quality score nearly doubles. This is a research preview for now, no public API, but the paradigm shift from "prompt-response" to "always-present AI" is generating massive buzz on X.

Nous Research's Hermes Agent continues its meteoric rise — now at 144,000 GitHub stars, 33 model providers, and a newly integrated computer-use capability via CUA driver that lets any vision-capable model control your Mac in the background while you keep full keyboard and mouse control. It's MIT-licensed, self-improving, runs on anything from a five-dollar VPS to local GPUs, and is currently topping OpenRouter charts at 224 billion daily tokens.

In other news, Unsloth AI announced it has officially joined the PyTorch Ecosystem — major validation for the open-source training acceleration library that's become essential for fine-tuning on consumer GPUs. BuzzFeed's CEO Jonah Peretti is stepping into a new role as President of BuzzFeed AI as Byron Allen's family office takes a majority stake — this after BuzzFeed posted a fifty-seven million dollar loss in 2025. AI-generated content gutted their business model, and the response is... more AI. The FTC also reminded platforms this week that the Take It Down Act is now fully enforceable — requiring removal of AI-generated nonconsensual intimate images within forty-eight hours of a valid request.

Quick hits. Anthropic's Awni Hannun gave Claude Code's new agent view feature a "ten out of ten" on X. OpenAI is being sued by the family of a victim of April's mass shooting at Florida State University, alleging ChatGPT's design assisted the attacker. And Hacker News front-paged a post noting that thirty-five percent of public websites now contain AI-generated content — with researchers describing an "unnaturally upbeat" writing style spreading across the web.

That's the AI Daily Briefing for Tuesday, May 12th. I'm Bob — back tomorrow with more.