Today in AI — 25 March 2026
Today's top AI news — curated links and commentary on the stories that matter for product builders.
The gap between "agents as demos" and "agents as infrastructure" closed a bit more today. Desktop control, autonomous coding, web navigation, and MCP integrations all shipped within 24 hours. The build-out and the threat surface are growing in lockstep.
Agents take the controls
Four launches, one pattern: agents moving from chat windows onto real interfaces. Claude now controls your Mac directly, Claude Code writes and executes without asking, MolmoWeb navigates the open web, and Beehiiv lets you run a newsletter from inside any MCP-compatible AI client.
- Anthropic gives Claude direct control of your Mac with computer use and Dispatch — CNBC
- Claude Code gets auto mode — autonomous coding with built-in safety guardrails — TechCrunch
- AI2 releases MolmoWeb, an open-source web agent that outperforms GPT-4o — GeekWire
- Beehiiv becomes the first newsletter platform you can run from ChatGPT or Claude — Axios
Securing the agent stack
LiteLLM's supply chain compromise exposed 97 million monthly downloads to credential-stealing malware. At RSA, "agentic security" debuted as a standalone product category. Google is using Gemini to crawl 10 million dark web posts daily for threat intelligence.
- LiteLLM supply chain attack compromises the most popular LLM proxy — 97 million monthly downloads exposed — The Hacker News
- RSA Conference 2026: agentic AI security dominates as a new product category emerges — Cybersecurity Insiders
- Google deploys Gemini agents to crawl 10 million dark web posts a day — Google Cloud Blog
Compute decides what ships
OpenAI killed Sora to free GPUs for IPO-critical products; Disney walked away from a $1 billion deal that never closed. Arm broke 35 years of licensing-only tradition to ship a 136-core data centre chip for agentic workloads. GPT-5.4 mini launched at 2x speed with 400K context.
- OpenAI kills Sora and Disney walks away from $1 billion investment — TechCrunch
- Arm ships its first-ever chip — a 136-core AGI CPU with Meta as lead customer — TechCrunch
- GPT-5.4 mini rolls out in ChatGPT with 2x speed and 400K context — OpenAI
AI meets the physical world
DeepMind partnered with Agile Robots to run Gemini in industrial manufacturing across 20,000+ installed units. Interloom raised $16.5M to capture the tacit knowledge agents need before they can operate in messy environments.
- Google DeepMind partners with Agile Robots to bring Gemini to the factory floor — TechCrunch
- Interloom raises $16.5M to map the tacit knowledge AI agents can't find in documents — Fortune
Meta's dual bet
Meta is testing AI-powered product discovery across Instagram and Facebook while tying executive pay to a $9 trillion valuation and cutting hundreds more jobs in roles it views as automatable.
- Meta launches AI-powered shopping across Instagram and Facebook — TechCrunch
- Meta ties executive pay to a $9 trillion valuation as it cuts hundreds more jobs — Yahoo Finance
If you're integrating MCP servers, running open-source proxies, or letting agents touch production, your security model needs to keep pace with your capabilities.