In partnership with
In Today’s Issue:
🛡️ Anthropic withholds its new Claude Mythos model from the public
💰 A group of OpenAI alumni launches a $100M venture fund to back practical AI startups
🚀 Jeff Bezos is quietly building "Project Prometheus"
💻 Z.ai drops a massive open-source model capable of working an uninterrupted
✨ And more AI goodness…
Dear Readers,
Anthropic just built an AI so good at hacking that they refused to release it, and that single decision might tell you more about where this industry is headed than any benchmark ever could. Today, we're unpacking Claude Mythos Preview, the frontier model that found a 27-year-old bug in one of the most secure operating systems on the planet, and why Anthropic's response was to lock it behind a coalition of tech giants rather than ship it to the public.
But cybersecurity isn't the only front moving fast: Jeff Bezos is quietly assembling a multi-billion-dollar AI empire called Project Prometheus that wants to leave chatbots behind and reshape aviation and engineering, while a group of OpenAI alumni just launched a $100M fund betting they know exactly which AI startups will survive the hype cycle.
On the open-source side, Z.ai's GLM-5.1 just proved that a 754-billion parameter model can work an uninterrupted 8-hour coding shift, autonomously building an entire Linux desktop from scratch, and that changes the math on what "AI-assisted development" actually means. Pour yourself a coffee, this one's dense.
All the best,
🚀 OpenAI Alumni Launch Bold AI Fund
A new VC fund called Zero Shot has raised its first $20M toward a $100M goal, founded by former OpenAI insiders including Andrew Mayne and Evan Morikawa. The team is already backing early-stage AI startups like Worktrace AI and Foundry Robotics, focusing on practical, high-impact applications while avoiding overhyped areas like “vibe coding” and digital twins.
Their edge: Deep insider knowledge of AI trends and strong builder networks, positioning them to spot what’s truly scalable in a fast-moving market where many bets may not pay off.
🚀 Bezos Expands Secret AI Ambitions
Jeff Bezos is rapidly scaling his stealth AI venture “Project Prometheus,” hiring Kyle Kosic, a former OpenAI and xAI leader, to build next-gen infrastructure. The company is aggressively recruiting top talent and aiming to develop AI that understands the physical world, targeting industries like aviation and engineering with massive proprietary datasets.
With plans to raise tens of billions and build a Berkshire-style AI investment powerhouse, Prometheus signals a shift from chatbots to real-world industrial transformation—where data, infrastructure, and talent wars are heating up fast.
🤖 Voice Mode Still Lacks Reasoning and tool Power
Despite being developed by OpenAI, ChatGPT’s Voice Mode still can’t reliably perform basic like timekeeping, according to CEO Sam Altman, who says it may take another year to fix. The issue highlights a broader weakness in AI systems like voice-mode handling time, numbers, and real-world tracking, sometimes even confidently giving wrong answers. It’s a reminder that even advanced AI can “hallucinate” simple tasks, reinforcing the need for human verification. And it’ll probably take some time until it can use tools.
“The dangers of getting this wrong are obvious, but if we get it right, there is a real opportunity to create a fundamentally more secure internet and world than we had before the advent of AI-powered cyber capabilities.”
— Dario Amodei about Claude Mythos
Our editor-in-chief Kim was a guest on a podcast again and talked about GPT-5.5 and the future of AGI.
Claude Mythos Rewrites Cybersecurity Rules
The Takeaway
👉 Anthropic withheld its most advanced model from public release because Mythos autonomously finds and exploits critical vulnerabilities in every major OS, browser, and widely used open-source software, including bugs that survived 27 years of human review.
👉 Project Glasswing gives 12 major tech companies and 40+ critical software organizations exclusive access to Mythos for defensive work, backed by $100M in usage credits, creating a first-mover advantage for defenders before similar capabilities proliferate.
👉 The gap between vulnerability disclosure and patching is now the most dangerous window in cybersecurity, as Mythos can turn a public CVE into a working exploit autonomously in hours, a task that previously took skilled researchers days or weeks.
👉 Organizations should immediately shorten patch cycles, invest in AI-driven security tooling with currently available models, and prepare vulnerability response plans for a world where thousands of critical bugs surface in weeks rather than years.
Anthropic just dropped a bombshell that could reshape how we think about AI and security forever. The company unveiled Claude Mythos Preview, a new frontier model so capable at finding and exploiting software vulnerabilities that Anthropic decided not to release it publicly. Instead, it launched Project Glasswing, a coalition including Apple, Microsoft, Google, AWS, NVIDIA, and CrowdStrike, giving these partners exclusive access to hunt bugs in critical infrastructure before attackers catch up.
The numbers are staggering. In just weeks, Mythos identified thousands of zero-day vulnerabilities across every major operating system and web browser. It found a 27-year-old bug in OpenBSD, an OS literally famous for being secure. It autonomously wrote a full remote code execution exploit for a 17-year-old FreeBSD flaw, no human help needed. Anthropic is committing up to $100 million in usage credits and $4 million in direct donations to open-source security organizations
The key tension however: the same capabilities that make Mythos a defensive superpower could be devastating in the wrong hands. Anthropic has privately warned government officials that Mythos makes large-scale cyberattacks significantly more likely this year. Can the industry patch fast enough to stay ahead of AI-powered offense?
Why it matters: Claude Mythos represents a fundamental shift where AI can discover and exploit vulnerabilities faster than humans ever could, threatening decades of cybersecurity equilibrium. The race between AI-powered offense and defense will define digital security for years to come.
Speak naturally. Send without fixing.
Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Not rough transcription you have to clean up. Actual polished text - ready for email, Slack, or any app.
Reid Hoffman sends 89% of his messages with zero edits using Flow. Millions of people worldwide have made it part of how they work, including teams at OpenAI, Vercel, and Clay.
Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.
Works on Mac, Windows, iPhone, and now Android - free and unlimited on Android during launch.
Perplexity doubled its revenue in a single quarter by launching “Computer,” a tool that shifts it from answering questions to actually completing tasks, unlocking much larger monetization beyond subscriptions.
Open Source AI Runs 8-Hour Shifts
What happens when you let an AI model work an entire 8-hour shift without interruption? Z.ai just answered that question with GLM-5.1, a 754-billion parameter open-source model that doesn't just write code, it obsessively refines it for hours on end. Released under the MIT License, GLM-5.1 tops SWE-Bench Pro at 58.4, beating GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro on complex software engineering tasks.
The real headline is endurance. In one test, GLM-5.1 was tasked with building a full Linux desktop environment as a web app, with zero starter code. Unlike previous models that produce a basic taskbar and declare the task complete, GLM-5.1 autonomously filled out a file browser, terminal, text editor, system monitor, and even functional games over 8 hours, iteratively polishing styling and fixing edge cases along the way. In another benchmark, it optimized a vector database across 600+ iterations and 6,000+ tool calls, reaching 6x the performance of any single-session run.
Z.ai's leader Lou wrote on X that GLM-5.1 can handle 1,700 autonomous steps, up from about 20 at the end of last year. That's not incremental improvement, that's a paradigm shift. If AI can work productively for 8 hours straight, what does the future software development cycle look like?
