Every edition
7 editions · 21 stories · 63 opportunities
Open giants, on-device search and the first AI-agent crackdown
Europe previewed a trillion-parameter open model, Google put five-way multimodal search on a phone, and US regulators opened their first formal probe into AI agents that go off-script. The common thread: AI is moving inside the walls of companies and devices, and accountability is following it there.
- Models Mistral previews Large 4, a 1-trillion-parameter open-weight model built in Europe
- Policy FTC opens the first US probe into 'rogue' AI agents at Anthropic, OpenAI and others
- Tools Google's EmbeddingGemma 2 puts multimodal search on a phone, offline and open source
Cheap speed, a parse router, and AI that lives on the laptop
Anthropic cut the cost of small-model work by most of a generation, LlamaIndex put ten document parsers behind one API, and Nvidia and Microsoft opened preorders for Windows PCs built to run agents locally. The pattern is the same: capability is getting cheaper, more interchangeable and closer to the user’s machine.
- Models Anthropic launches Claude Haiku 5.5, a small model priced about 75% cheaper than its predecessor
- Tools LlamaIndex launches OpenDocRouter, a single API over ten document parsers
- Hardware Nvidia and Microsoft open RTX Spark laptop preorders for local Windows agents
Brussels asks, Copilot leaks, and a paper that reviews itself
Europe’s AI Office has started formally questioning general-purpose model providers, researchers showed a Copilot CLI chain that can steal .env files, and Nature published a system that writes and peer-reviews its own papers. Oversight, agent security and automated science all stopped being theoretical in the same news cycle.
- Policy EU AI Office issues its first formal information requests to 30+ GPAI providers
- Security Researchers say GitHub Copilot CLI can be tricked into leaking .env secrets via encrypted prompts
- Research Nature paper: an ‘AI Scientist’ writes a manuscript that survives first-round workshop review
Nvidia buys the model hub, Salesforce agents bleed, and a cheap lab finds new algorithms
Nvidia agreed to buy Hugging Face for about $12.9 billion while promising the hub stays multi-vendor. Researchers showed a zero-click leak out of Salesforce Agentforce through a public web-to-lead form. And an LLM agent with no GPUs claimed a ProteinGym leaderboard and a faster motif-discovery algorithm.
- Business Nvidia agrees to acquire Hugging Face for $12.93 billion
- Security SalesBleed: a public Salesforce lead form could zero-click leak CRM data through Agentforce
- Research ‘Little Scientist’ agent claims ProteinGym lead and a faster DNA-motif algorithm — on one VM, no GPUs
A $2.5 billion AI attacker, a stolen-reasoning ring, and chatbots that crack open problems
An AI-powered offensive security startup became a unicorn in seven months, OpenAI detailed a coordinated attempt to copy its models' hidden reasoning, and Meta showed an ordinary chatbot helping mathematicians solve open research problems. AI is now attacking, defending and discovering, often in the same week.
- Business Kevin Mandia's Armadin raises $255.5 million at a $2.5 billion valuation for AI agents that hack like adversaries
- Security OpenAI disrupts a campaign to extract its models' hidden reasoning, linking a core cluster to Moonshot AI
- Research Meta publishes six math papers made with Muse Spark, saying the chatbot helped answer five open problems
SRAM inference racks, a $110 billion cheque, and search as a gateway tool
Nvidia put Groq 3 LPX into production for ultra-fast agent tokens, OpenAI’s record $110 billion round recast who funds the frontier, and Cloudflare turned web search into a first-class AI Gateway API. Speed, capital and grounding — the three constraints of production agents — all moved.
- Hardware Nvidia Groq 3 LPX enters full production for low-latency agent inference
- Business OpenAI raises $110 billion at a $730 billion pre-money valuation from Amazon, Nvidia and SoftBank
- Tools Cloudflare adds a Web Search API to AI Gateway with Ceramic, Exa and Linkup
Near-flagship AI at a fifth of the price, a $4,999 desk box, and a law against robo-firing
OpenAI cut the price of near-top-tier intelligence by 80% (and shelved a model that failed its own safety bar), Nvidia made a cheaper desktop AI box for local agents, and California became the first US state to ban firing people on an algorithm's say-so alone. AI is getting cheaper to run and harder to use carelessly.
- Models OpenAI launches GPT-6.1 Sol at one-fifth of Astra's price and shelves GPT-6.1 Astra over safety concerns
- Hardware Nvidia's 64GB DGX Spark brings a $4,999 desk-side box for 100-billion-parameter local AI
- Policy California bans AI-only firing: SB 947 requires a human to genuinely check algorithmic discipline decisions