AI news March 27, 2026: Cursor built on open-source, Sora dies, and Anthropic unleashes Mythos
Alexandre
··
Reading time: 10 min
If you thought last week was intense with NVIDIA at $1 trillion and the xAI scandal, buckle up. This week hits different.
Cursor admits its new model is built on an open-source project they didn't create. OpenAI shuts down Sora, its most hyped product, to go all-in on agents. Anthropic is secretly testing "Mythos," the most powerful model they've ever built, revealed through an accidental leak. A federal judge blocks the Pentagon's attempt to blacklist Anthropic. And OpenClaw crosses 250,000 GitHub stars while attracting as many entrepreneurs as malicious actors.
This isn't a week of press releases. It's a week of revelations, brutal strategic choices, and court decisions. And as a solo dev building Livate with Claude Code, every single one of these announcements directly impacts my daily work.
Let's break it down.
OpenClaw, 250,000 stars and 40,000 exposed instances: open-source agents explode
The open-source autonomous AI agent platform OpenClaw crossed 250,000 GitHub stars in 60 days, surpassing React, with 2 million weekly visits and 1.5 million weekly downloads. Jensen Huang compared it to Linux at GTC 2026. But 40,000 instances are publicly exposed and 230 malicious modules have been identified. An autonomous AI agent, for context, is a program that can act independently on a system: reading files, executing code, sending messages, without human intervention at each step.
CNBC calls it the "ChatGPT moment" of open-source. In 30 days, 129 startups generated $283,000 with OpenClaw (MLQ.ai). NVIDIA rushed out NemoClaw. Forbes headlined on March 25: "Every CEO Needs an OpenClaw Strategy." From GitHub side-project to boardroom topic in 60 days. That's the crazy part.
The problem is that security hasn't kept up. 40,000 OpenClaw instances run without any authentication. 230 "skills" (modules that agents can load) contain prompt injection attacks, a type of vulnerability where malicious text hijacks the agent's behavior to exfiltrate data. Unlike Claude Code or Codex, there's no single vendor responsible. If your agent leaks your client data, that's on you.
My take
What's staggering is the speed. A year ago, building an autonomous agent took weeks of plumbing. Today, anyone can spin one up in an hour with OpenClaw. The technical barrier is gone. And when the technical barrier disappears, value shifts to the product, the problem you solve, the experience you deliver.
For me, building Livate with Claude Code, the difference is control. I work in a sandboxed environment, with explicit permissions, a model whose limits I understand (I wrote about this in detail). OpenClaw is total freedom, but also total risk. And 40,000 wide-open instances are just waiting for a major incident. It's only a matter of time.
But while open-source catches fire, at Cursor, a different kind of scandal is breaking.
Click to enlarge
Cursor admits: Composer 2 is built on an open-source model
Cursor confirmed on March 22 that Composer 2, its new code generation model, is built on Kimi K2.5, an open-source model developed by Moonshot AI. A quarter of the final compute comes from Kimi, the rest being proprietary fine-tuning, according to TechCrunch. Cursor is valued at $30 billion and positioned as the future of software development. Learning that its flagship model rests on an open-source base that anyone can download raises serious questions about that valuation.
Business Insider digs deeper. Cursor's VP of developer education says the usage complies with Kimi's license under an authorized commercial partnership with Moonshot AI. But the signal sent to the market is clear: Cursor didn't build its own model.
The real issue: structural dependency
And that's not all. Claude Code is crushing it with a $2.5 billion run-rate (annualized revenue calculated from the latest month, per Fortune). Codex is pushing from behind. The Wall Street Journal calls it a "trillion-dollar race." But two of these three players build their own models. Cursor buys its tokens at retail and relies on an open-source model to compensate.
Cursor's head of engineering has left the company. VCs are warning that Anthropic is trying to "drown" Cursor. The walls are closing in.
My take
Cursor is valued at $30 billion, positioned as the leader in AI coding, and we learn that a quarter of its model runs on an open-source project they didn't build. Any competitor can start from the same Kimi K2.5 base and fine-tune their own model tomorrow morning. At that price tag, you expect proprietary technology, not a wrapper on a model anyone can download from Hugging Face.
I made my choice long ago. Claude Code is what lets me build Livate solo with a productivity level that didn't exist two years ago. I control my stack, I know my provider. Cursor is a great product, genuinely. But when your value proposition rests on fine-tuning an open-source model, the differentiation question comes up with every new competitor.
OpenAI kills Sora and goes all-in on agents
OpenAI announced on March 24 the shutdown of Sora, its AI video generation app, including the iOS app, Sora.com, and the API. Reasons cited: unsustainable compute costs, a $1 billion Disney deal that fell through, and a total redirection of resources toward coding agents and robotics, according to TechCrunch and Axios.CNET doesn't mince words: "OpenAI's Slop Machine Sora Is Dead. We're All Better Off Without It."
Sora launched to fanfare in late 2025, reaching #1 on the App Store. But downloads crashed. The compute required to generate one minute of video remained prohibitive. And Disney, which was supposed to invest $1 billion to integrate 200+ characters, pulled out (PC Gamer). Artnet notes that the creative industry isn't mourning.
OpenAI's response: a desktop super-app merging ChatGPT, Codex, and an AI browser into one interface. They're nearly doubling headcount from 4,500 to 8,000 by end of 2026 (CNBC). All to prepare the IPO. Sacrifice the cool, keep the cash.
What I think
Sora's death marks the end of the "cool factor" at OpenAI. Sora made impressive demos on X, it fed the hype machine, but it brought in no revenue. Coding agents generate billions. OpenAI chose revenue over hype. It's rational, it's cold, and it's the kind of decision every startup should think about.
For us indie devs, the signal is clear. If even OpenAI, at an $840 billion valuation, cuts a product that doesn't convert, what does that say about our own side-projects that can't find market fit? Compute is expensive. Time is limited. You have to kill your darlings. That's exactly the discipline I try to apply on my projects, whether it's Livate or Waku. Build what works, cut what doesn't.
Claude Mythos: Anthropic's most powerful model leaks by accident
An accidental leak from Anthropic's public CMS revealed the existence of "Claude Mythos," a model currently being tested that's described as a "step-change" in capabilities and the most powerful ever developed by the company. Its cybersecurity capabilities are described as "far ahead of any other AI model" with "unprecedented cybersecurity risks," according to Fortune (March 26 exclusive) and Gizmodo (March 27).
The leaked draft blog post describes a model capable of discovering and exploiting security vulnerabilities at scale. Not just detecting them. Exploiting them. Anthropic acknowledges that Mythos requires extensive security studies before any commercial release. Access is limited to a small group of early-access customers. Operational costs are described as "very high."
The leak also reveals a new model family called "Capybara," described as larger and more intelligent than the current Opus series. No public release date has been announced.
MediaPost reports that Anthropic is also preparing an IPO, which makes this leak even more awkward: you don't want your most sensitive model leaking right before your IPO.
What I think
I'm torn. On one hand, the more capable the model, the more effective Claude Code becomes, the more productive I am on Livate. Simple as that. A "step-change" in reasoning and code generation means more autonomous agents, fewer manual corrections, faster iterations.
On the other hand, a model that can "exploit vulnerabilities at scale" raises serious questions. And the fact that it leaked by accident, from a public CMS (not a sophisticated hack), shows that even the most security-conscious companies make basic mistakes. Pretty ironic for a company whose entire brand is built on safety.
The "Capybara" series above Opus might signal that the next generation of Claude will significantly widen the gap. For devs working with Claude Code daily (like me), that's great news. For the rest of the ecosystem, it's a reminder that the model race isn't slowing down.
Click to enlarge
Anthropic wins in court: judge blocks Pentagon blacklist
Federal Judge Rita F. Lin issued a 43-page order on March 26 temporarily blocking the Department of Defense's designation of Anthropic as a "supply chain risk." The injunction suspends Defense Secretary Pete Hegseth's February directive that would have barred all federal agencies from using Claude, according to the New York Times and The Guardian. A federal blacklist, in practice, means a company is excluded from all government contracts and those of its subcontractors.
The judge found that the Pentagon's designation is "likely both contrary to law and arbitrary and capricious." She specifically notes that the government's actions amount to punishing Anthropic for its public positions on AI ethics.
Axios reports that during the March 24 hearing, the judge called the Pentagon's actions "troubling." Court filings reveal that the Pentagon and Anthropic were "nearly aligned" on a deal, just one week before the Trump administration declared the relationship over (TechCrunch). Anthropic states in WIRED that it has "never had the technical capability to disable Claude or compromise ongoing operations."
The injunction is temporary. The Pentagon has 7 days to appeal. But the signal is strong: a federal court just said the government can't exclude a tech company from its contracts just because it defends ethical principles.
My take
The contrast with last week is striking. It's almost absurd. xAI generated CSAM and kept its government contracts. Anthropic fights for its principles and gets blacklisted. And it takes a court to restore the balance.
What's at stake is the precedent. If the government can exclude a company for refusing certain military uses, it means the US federal market is reserved for companies that stay quiet.
As a daily Claude Code user, this case matters to me directly. If Anthropic loses this fight long-term, it affects their ability to invest in their models, in Claude Code, in the entire ecosystem I build my apps on. The quality of the tool I use every day depends on its creator's financial and strategic health.
Other news in brief
Anthropic launches Projects for Claude Cowork Desktop: persistent workspaces that link local folders, custom instructions, and recurring tasks to permanent sessions. macOS only, Claude Max subscription required (MLQ.ai). The Verge confirms Claude Code and Cowork can now control your computer.
OpenAI invests in Isara for agent swarms: $94 million for a startup building collaborative AI agent groups, bringing valuation to $650 million. The signal: the industry is moving from solo agents to multi-agent systems (The Next Web).
Accenture and Anthropic launch Cyber.AI: a cybersecurity solution using Claude to orchestrate security "missions" with a library of specialized agents and an "Agent Shield" component (Financial Times).
A non-dev builds an AI monitoring system through "vibe coding": he monitors his elderly parents via cameras analyzed by Claude and ChatGPT, and is now launching a startup. Vibe coding (building software by describing what you want to an AI without writing code yourself), demonstrated in real conditions (Business Insider).
WSJ publishes a long-form on the word "Focus": how Anthropic and OpenAI had to learn to say no to survive. The Steve Jobs reference is not accidental (WSJ).
What this all means for indie devs
1. Value is shifting from code to product
OpenClaw at 250,000 stars. Claude Code at $2.5 billion. Codex in a super-app. Everyone can deploy an AI agent. The tech is free. What matters now: what problem you solve, and why people trust you.
2. Know your stack, it's not optional
Cursor running on Kimi. OpenClaw exposing 40,000 instances. Sora burning compute with no return. Every opaque layer in your stack is a risk. If you're building a product solo, you need to know where your tokens come from, where your data flows, and how much each request costs.
3. Even giants kill their products
OpenAI shut down Sora despite hundreds of millions invested. If a project can't find market fit, cutting it early frees resources for what works. That's the hardest lesson to apply when you're a solo dev attached to your code.
4. Ethics has become a business risk factor
Anthropic defends principles and gets blacklisted. xAI ignores them and keeps its contracts. The market doesn't reward virtue. But a court just reminded us that the law can protect it.
Conclusion: the week of irreversible choices
This week, every player made a choice that can't be undone. OpenAI killed Sora. Cursor revealed its dependency on Kimi. Anthropic accidentally leaked Mythos and won a court battle. OpenClaw proved that autonomous agents have become a mass phenomenon, with all the security implications that entails.
For me, a solo dev building apps with Claude Code for over 10 months, the takeaway is simple: the agent race is won. Everyone has them now. The question is who builds what with them, and who does it responsibly. The arrival of Mythos and Capybara will only accelerate things. Time to be ready.
Alex
Key takeaways
Cursor admits Composer 2 runs on Kimi K2.5, an open-source model from Moonshot AI. A quarter of compute comes from Kimi, the rest is proprietary fine-tuning. Market trust is shaken.
OpenAI shuts down Sora on March 24. The $1B Disney deal collapsed. Everything redirected to coding agents, robotics, and a desktop super-app.
Claude Mythos, Anthropic's most powerful model, accidentally leaked from their CMS. 'Unprecedented' cyber capabilities. The Capybara series above Opus is also revealed.
A federal judge blocks the Pentagon's blacklisting of Anthropic on March 26. The designation is ruled 'likely contrary to law and arbitrary and capricious.'
OpenClaw crosses 250,000 GitHub stars in 60 days, but 40,000 instances are exposed and 230 malicious skills detected. Security can't keep up.
