Today in AI — 7 May 2026
Today's top AI news — curated links and commentary on the stories that matter for product builders.
Anthropic, Google, and OpenAI are all building toward agents that never stop running. Two independent security audits suggest nobody has figured out how to make that safe yet.
Agents that never clock out
Three companies, one thesis: agents should keep working when humans stop watching. Anthropic ships dreaming and Routines; Google tests the same persistent-agent idea with Remy inside Gemini.
- Anthropic teaches Claude agents to 'dream' between tasks — SiliconANGLE
- Claude Code launches Routines so AI agents can ship code while you sleep — Anthropic
- Google tests Remy, a 24/7 personal AI agent that acts on your behalf inside Gemini — Artificial Intelligence News
Powering the always-on agents
Always-on agents need always-on compute. Anthropic signed a SpaceX deal for 300+ MW, RadixArk raised $100M to commercialise SGLang, and Google shipped MTP drafters for Gemma 4 with up to 3x inference speedup.
- Anthropic signs SpaceX deal for 300MW of Colossus compute and explores space-based AI — CNBC
- RadixArk raises $100M seed to scale SGLang, the open-source engine behind trillions of daily AI tokens — Tech Funding News
- Google releases MTP drafters for Gemma 4 with up to 3x inference speedup — Google Blog
Agent security is still broken
OX Security found MCP's STDIO transport executes OS commands with no sanitisation, exposing 200,000 servers. A separate audit found open-source agent skills riddled with invisible supply-chain flaws. The ecosystem is growing faster than its trust model.
- 200,000 MCP servers exposed by command execution flaw that Anthropic calls a feature — VentureBeat
- One command turns any open-source repo into an AI agent backdoor — VentureBeat
Google's search and simulation bets
Two contrasting plays: Google is folding human forum knowledge into AI search results, while DeepMind invested in the EVE Online studio to drop AI into a player-driven economy and observe what happens.
- Google adds Reddit and forum quotes to AI search results — TechCrunch
- Google DeepMind invests in EVE Online studio to test AI in a player-driven economy — 9to5Google
OpenAI ads and the Musk trial
OpenAI is sprinting toward ad revenue, opening self-serve ads to all US businesses. Prediction markets give Musk just 40% odds of winning his suit after Brockman's testimony.
- OpenAI opens ChatGPT Ads Manager to all US businesses — OpenAI
- Musk's odds of winning OpenAI trial drop to 40% as Brockman wraps testimony — CNBC
The tooling for persistent agents is arriving faster than the safeguards; builders should assume that whatever they ship today will face retroactive security standards tomorrow.