Today's VB Daily read time: ~3 minutes. Good morning. Manufact raised $6.3 million to build the infrastructure layer for MCP, the protocol Anthropic introduced in late 2024 that has become the standard way AI agents connect to software: 10,000+ active public MCP servers; 7 million downloads per month; ChatGPT, Cursor, Google Gemini and VS Code all support it; and Anthropic donated it to the Linux Foundation in December. Manufact's open-source library mcp-use — "connect any MCP to any LLM in six lines of code" — has 5 million downloads, 9,000 GitHub stars, and is used by NASA, Nvidia, and SAP. The company has three employees.
Further down: An autonomous agent breached McKinsey's AI platform through a SQL injection nobody caught for two years, Nvidia committed $26B to open-weight models, and Notion's AI lead on why frontier labs have no incentive to build affordable mid-tier models
|
|
|
Manufact raises $6.3M as MCP becomes the ‘USB-C for AI’ powering ChatGPT and Claude apps |
|
|
A message from Microsoft and NVIDIA |
Know before you go: What’s new with Microsoft at NVIDIA GTC They say it all starts here at NVIDIA GTC. Where will you start? Explore ways to engage with Microsoft at the leading AI conference of the year. Read the Azure Tech Community Blog for all the details on Microsoft-led talks and panels, in-booth sessions, hands-on labs, demos, and more.
Make the most of every day at GTC when you prepare ahead with activities aligned to your AI goals. See what’s coming and mark your calendars for what matters most to your business. We can’t wait to see you there. |
|
|
Perplexity takes its ‘Computer’ AI agent into the enterprise, taking aim at Microsoft and Salesforce |
Who's actually in charge of your AI stack? VentureBeat is surveying enterprise technology and engineering leaders to find out what's really happening on the ground: who governs AI behavior across complex, multi-platform environments, and what's getting in the way. |
An autonomous AI agent breached McKinsey's internal AI platform in two hours A CodeWall.ai autonomous agent breached McKinsey's internal AI platform — Lilli, used by 43,000+ employees — through SQL injection. It accessed 46.5 million chat messages, 728,000 files, 57,000 employee accounts, and 3.68 million RAG document chunks. The vulnerability sat in production for over two years, and OWASP ZAP missed it. Read more at CodeWall.ai OpenAI releases IH-Challenge, an RL training dataset that teaches models to follow instruction hierarchy OpenAI released IH-Challenge, an RL training dataset that teaches models to follow instruction hierarchy: system > developer > user > tool. Fine-tuning GPT-5-Mini on it produced +10% average improvement across 16 benchmarks and dropped unsafe behavior from 6.6% to 0.7%. Read more at OpenAI Nvidia commits $26B to open-weight models as U.S. alternative to DeepSeek The investment provides a U.S.-made alternative as startups and researchers increasingly build on Chinese open models from DeepSeek, Alibaba, and Moonshot AI. The company released Nemotron 3 Super (128B) and recently finished pretraining a 550B model. Read more at Wired |
|
|
Beyond the Pilot Podcast: Episode 8 LangChain told employees they cannot install OpenClaw on company laptops due to "massive security risk" — yet this unhinged approach is exactly what makes it work. Harrison Chase unpacks why OpenClaw succeeds where AutoGPT failed, and why context engineering, not just smarter models, separates demo agents from production-ready systems. Watch Episode 8 |
For more on these stories and other key updates, head over to VentureBeat.com for our full coverage. - Bryson Masse (@Bryson_M) For U.S.-based readers, don't miss VB on Google: Add us to your trusted feeds. Click here. |
|
|
|
