• The AI Furnace
  • Posts
  • AI Furnace Newsroom: ​​Anthropic’s Leaked “Mythos” Model, Cursor Takes Swing at Anthropic and OpenAI, Claude Can Now Operate Your Computer

AI Furnace Newsroom: ​​Anthropic’s Leaked “Mythos” Model, Cursor Takes Swing at Anthropic and OpenAI, Claude Can Now Operate Your Computer

The AI Furnace weekly newsletter. Your destination for the latest news, innovations, opportunities, and product launches in AI.

Author

AI Furnace
March 31, 2026

In partnership with

Welcome to this week’s AI Furnace Newsroom

In today’s insights we cover:

  1. Anthropic’s Leaked “Mythos” Model Shows Where the Cyber Risk is Headed

  2. Cursor Takes Swing at Anthropic and OpenAI with Composer 2 Launch, but Not Without Controversy

  3. Claude Can Now Operate Your Computer

  4. The AI Jobs Slash Continues to Dominate Headlines, but the Data is Getting Messier 

  5. OpenAI Kills Sora, and that Says a Lot About Where the Money is Going

  6. Mistral’s New Voice Model Can Run on Edge Devices and Your Smartphone

Read time: 5 mins

💡 Furnace Insights

Anthropic

Anthropic accidentally exposed draft materials for an unreleased model called Claude Mythos, which the company later confirmed is real and described as “the most capable” system it has built so far.

The leak may have been a marketing stunt cooked up with Fortune magazine, who supposedly accidentally found the data online and reported it to Anthropic, but hey - we are in the wild west of AI and the AI labs are competing for mindshare. The leaked material said Mythos belongs to a new Capybara tier above Opus, and flagged unusually strong performance in reasoning, coding, and especially cybersecurity, with Anthropic saying it is being tested only with a small group of early-access users for now.  

The leak itself suggests that in the next generation of models cyber capability is going to become a first-order issue as hackers utilize AI on the offensive, and also that leaks could be the way to market new AI products.

Cursor

While there are concerns that Anthropic and OpenAI are taking over AI coding budgets, Cursor has still grown fast as a business with annualized revenue reportedly above $2 billion and a valuation around $29.3 billion. Last week Cursor launched Composer 2, a new coding model/AI agent that can carry out coding tasks on a user’s behalf. But the story quickly shifted when developers found identifiers linking it to Moonshot AI’s Kimi family - a Chinese model, under the hood. 

Cursor later acknowledged that Composer 2 was built on Kimi, saying it added its own reinforcement learning and other training on top, and admitted the missing attribution was a mistake. Cursor is trying to compete with giants like Anthropic and OpenAI, so this reputational hit wasn’t good timing to say the least.

Anthropic

Anthropic has given Claude the ability to directly control a user’s Mac desktop inside Claude Cowork and Claude Code, with the feature live as a research preview for macOS. Anthropic says that this Claude feature first tries to do tasks via app integrations, and then falls back to direct desktop control when no suitable interface exists. It also launched Dispatch, which lets users send tasks to their own machine remotely. The move builds on Anthropic’s recent acquisition of Vercept, a startup focused on computer-use agents.  

The desktop is becoming the real battleground because that is where white-collar work actually happens. OpenClaw got attention by promising a similar kind of AI that “uses your computer for you”, and Anthropic is trying to package the same idea with tighter controls and a more enterprise-friendly posture.

AI Market

A recent Gartner forecast relating to CFO hiring says around half of companies that cut customer-service headcount because of AI will rehire for similar roles by 2027, often under different titles. At the same time, companies are still openly linking workforce reductions to AI: Crypto.com said it would cut 12% of staff as it integrates AI more broadly, while reports say HSBC is weighing large back- and middle-office cuts as it expands generative AI across the bank.  

That mix tells you something important. AI is already changing hiring and org design, but a lot of companies are still in the phase where they over-cut, discover the limits of automation, and then rebuild around a different human-machine mix. So the near-term labor story is probably not “humans gone”. It is churn: fewer traditional roles, more re-organizations, and a lot of uncertainty about which tasks really hold up once AI leaves the demo and hits production. 

OpenAI

OpenAI is discontinuing Sora, its video generation model which only launched six months ago but which achieved some impressive feats such as 1 million downloads in the first 24 hours of launch. But it also consumed 30% of OpenAI’s entire compute capacity on a given day, faced heavy moderation issues, copyright controversy and weak retention. It seems OpenAI is choosing focus over spectacle, as it consolidates its various products and sharpens itself to compete with Anthropic for the enterprise.

Mistral

Mistral has released Voxtral TTS, a text-to-speech model that supports nine languages, offers zero-shot voice cloning, and is designed to run efficiently enough for edge devices like phones and laptops. Mistral is positioning it as part of a broader multimodal platform, and outside coverage has framed it as a direct challenge to companies like ElevenLabs, OpenAI, and Deepgram.  

Voice AI is heating up with production deployments growing 340% year over year and Gartner predicting conversational AI will cut contact center labor costs by $80 billion in 2026 alone. Mistral is throwing its hat in the ring - it is no longer competing in a “make the voice sound nicer” race. Enterprise buyers care about latency, privacy, deployment flexibility, and primarily cost. By going open-weight and edge-friendly, Mistral is betting that this is the future of voice in the enterprise. 

📈 AI Venture Deals of the Week

  • Deeplify, an AI workflow automation platform, raised its $2M Pre-seed.

  • AI&, an enterprise AI agent platform, raised its $50M Seed.

  • eMed, an AI-enabled telehealth platform, raised its $200M Series A.

  • Qualified Health, an AI platform for health systems, raised its $125M Series B.

  • Doctronic, an AI-powered virtual doctor platform, raised its $40M Series B.

  • Assiduus Global, an AI e-commerce growth accelerator, raised its $25M Pre-series B.

  • Granola, an AI meeting notes platform, raised its $125M Series C.

⚒️ New AI Product Launches You Don’t Want to Miss 

Upcoming Events 📅

Interested in meeting the Who’s Who of AI? Join the AI Hot 100 Summit in NYC

The AI Hot 100 Summit is back for its third edition on May 6-7th, 2026 in New York City. The only AI Summit built for real connections where AI Visionaries meet Industry Leaders. Join over 500 AI Executive Leaders, Founders and Investors for a jam packed 2 days of enterprise AI lightning talks, innovation showcases, and curated networking.

Get ready to meet and greet speakers and attendees from OpenAI, NVIDIA, Haleon plc, Legora, Anthropic, Cursor, Eleven Labs, Adobe, Walmart, Insight Partners & more.

🗓️ May 6-7, 2026

📍New York City

📢 Want to partner? Reach out to get your brand in front of 25k+ AI executives, entrepreneurs, researchers, investors, and AI leaders.