What's moving in AI

AI safety is becoming productized through vulnerabilities, watermarks, and policy turbulence

Anthropic’s Project Glasswing found more than 10,000 high- or critical-severity vulnerabilities, while Google DeepMind expanded SynthID detection through Gemini and Search. Ars Technica’s canceled AI-safety EO report shows the split between useful safety infrastructure and political theater.

Multimodal generation is moving closer to everyday content workflows

The Verge covered Google’s Gemini Omni as an anything-to-anything deepfake/video generation experience. The direction is mixed-input creation where image, video, reference, text, and world context become editable ingredients.

Enterprise AI keeps becoming branded experience infrastructure

TechCrunch reported that Ferrari is using IBM’s AI to create F1 superfans. It is a useful enterprise AI example because the value comes from deepening a domain-specific user journey rather than shipping a generic productivity feature.

Model labs are becoming agent-product companies

The model alone is no longer the product: Cursor is exposing Composer 2.5 through an SDK while Dropbox Nova and Codex analytics show labs and dev-tool vendors building full agent harnesses. The moat is shifting toward UI, memory, sandboxing, permissions, cost controls, evals, and integration depth.

Google is turning search into an action surface — with messy side effects

Google is pushing AI Mode, AI-generated ads, and broader Gemini/Search integrations, but The Verge highlighted AI Overview failures where the system can drift into chatbot-like behavior and appear to disregard the query. The distribution upside comes with trust and attribution risks.

The agent economy is becoming a compute-and-distribution knife fight

Cursor revenue momentum, Anthropic compute-deal rumors, and DeepSeek's permanent DeepSeek-V4-Pro discount all point to the same pressure: agents need cheaper inference and massive distribution. Cheaper tokens make longer-running workflows more practical.