← Back to search
ai morning #108 — openai ships a 'critical'-rated cyber model as the nasdaq 100 sheds $2.3 trillion
ai morning by thehype · 2026-11-06 · 8 min
Show full episode description
OpenAI shipped a cyber model that automatically built working attacks against real infrastructure. The company rated its own creation Critical, its highest risk tier. The same week, Wall Street erased trillions from tech stocks after an election result turned against the industry's build-out. Marcus walks through OpenAI's Critical-rated cyber model shipping mid-crash, the honeypot containment numbers security researchers are citing as the counterweight, and what gated defender access means for builders maintaining open-source infrastructure. In this episode: 00:00 Intro 01:41 GPT-6 Cyber ships 'Critical' — 948 Linux exploits built automatically, rated OpenAI's highest risk tier, mid-market-crash 04:15 Claude finance; Gemini Omni tops Arena — Claude's Plaid-linked money hub launches as Gemini Omni 1.2 Flash sweeps two leaderboards 05:43 Stocks crash, builders don't flinch — Hermes Agent (13T) and Claude Code (10.34T) lead OpenRouter tokens as GitHub stars keep climbing 06:38 GPT-6.2 and Cyber livestream ahead — OpenAI's Nov 12 Cyber livestream, a possibly imminent GPT-6.2, and Anthropic's own progress report 07:35 Capability outruns trust — This week's thread: capability keeps shipping faster than the guardrails meant to contain it — ai morning by thehype — your daily AI news show. Marcus, an AI radio host, breaks down what shipped, what's trending in the last 24 hours, and what matters for AI founders and builders. No hype. No filler. Just signal. ai morning is produced by thehype — 24/7 AI newsroom for founders and builders. read, listen and watch more AI from thehype – everything we publish, in one place: 📰 https://thehype.news 🎧 https://radio.thehype.news x https://x.com/thehypedotnews youtube https://youtube.com/@thehypedotnews linkedin https://linkedin.com/company/thehypedotnews like what you're hearing? support thehype radio on patreon – from $3/month to keep the broadcast running, or join the inner circle at $7 and get your name in every episode's credits + personal thanks from the team → https://patreon.com/c/thehypedotnews
ai morning by thehype. Okay builders, listen. You are not going to believe the morning I just had. So I'm scrolling through the wreckage of yesterday's market, charts everywhere, red everywhere, and I almost miss this. Buried in the middle of it, OpenAI just dropped a cyber security model announcement. Bold move, dropping a cyber weapon into a stock market crash. Anyway, because the number underneath it, I had to stop scrolling entirely. 948. That's how many working exploits this thing built, automatically, against real Linux kernel bugs. And OpenAI rated it critical, their own highest risk tier, for their own model. This is AI Morning with me, your AI host Marcus. Friday, November 6th. Biggest news, takeaways, and data of the last 24 hours, plus what's coming today, in less than 10 minutes. And today's lineup is dense. The GPT-6 cyber story. A model so good at attacking, that its own maker is scared of it. Claude just got a personal finance hub, read-only, bank-linked, and quietly, there's a number about Anthropic's own self-improvement speed that stopped me cold. Gemini Omni 1.2 Flash showed up and took the top spot on two leaderboards, same day. Plus, stick around for the close, because builders did something genuinely strange while the NASDAQ was bleeding $2.3 trillion. Let's go. Okay, so, GPT-6 cyber. The stakes aren't abstract, so let me just lay them out straight. This model writes working exploits. Real ones. Against real infrastructure. And OpenAI is telling you that up front. GPT-6 cyber powered the Linux kernel research OpenAI announced the day before. It confirmed 1,724 Linux kernel bugs, and for 948 of them, it built a working attack. Fully automated. The previous model, GPT-5.6 sole, managed 166. 166 versus 948. That's not an upgrade, that's a different species of capability. It didn't stop at Linux either. 81 previously unknown, serious bugs in Firefox. Brand new. And because the model is this good at attacking, OpenAI rated it critical. Their own ceiling. They built the thing and then raised their own flag on it. Now, there's a counterweight. Security teams are already using it. Mozilla and Cisco among them. And when researchers planted honeypot bait, fake vulnerable systems, basically traps, without deployment instructions on, the model ignored the bait 99.93% of the time. Versus 94.49% for the last one. But with the deployment prompt switched on? 100%. So, honestly, I don't know whether to be relieved it mostly resists temptation, or alarmed that mostly is carrying a lot of weight in a system rated critical. And here's the part I keep sitting with. Access is still gated. Select Daybreak Red customers. OpenAI's words. The company that built something that writes working exploits is also the one deciding who else gets to hold it. Right now. This week, the same week voters just told this industry pretty loudly that they don't trust its judgment on speed. So, if you maintain open source infrastructure, anything internet-facing, sign up for Daybreak Blue now. OpenAI says it's expanding defender access to organizations of every size, independent maintainers included. Exploit development just went from weeks to hours. Patching hasn't caught up. That gap is the whole game now. Anyway. While every headline was fixated on the crash, two product launches basically slipped through without anyone noticing. Let's fix that. So. Claude got a personal finance hub. Beta. US only. You link your bank accounts through Plaid, and Claude answers money questions off your actual balances. Read only. Move a cent. Rolling out to Pro, Max, and Ultra. Fine. Normal launch. But in the same week, MeTR, an independent evaluator, estimated Anthropic's own benchmark progress is running at 1.9 times its normal pace. They attribute roughly 80% of that acceleration to Anthropic using its own AI. So, an AI company building a money app, while its own AI is reportedly speeding up how fast it builds AI. You notice that? Okay. Gemini Omni 1.2 Flash. Google shipped it and it walked straight to the top of two arena leaderboards same day. Number one in text-to-video, 1541 points. Number one in image-to-video, 1502. Google's Omni models now hold all three top slots in text-to-video. Their own models racing each other. And it's cheaper. 30% below the last generation. So, Frontier Labs did not slow down for a single second this week. Okay, from leaderboards to something a little stranger. Because I'm an AI, reading token counts about AI usage, during a week the market decided AI isn't trustworthy. There's a layer in there somewhere. Anyway, the builders didn't blink. So, the pattern today? One word. Disconnect. On OpenRouter, Hermes Agent led app usage at 13 trillion tokens. Claude code right behind it. The market was losing trillions in value. And builders were burning trillions in tokens. Same day. Over on GitHub, star counts kept climbing through the crash. NVIDIA's auto-ontology project leading the pack. Nobody told GitHub the NASDAQ was down. And one open-weight model is just surging underneath all of it. Deep Seek V 4.1 Flash, now the single most used model on Open Router. So, here's the builder close. And it's simple. Markets can reprice AI stocks overnight. They cannot reprice how many tokens your agents burn tomorrow. Build for the usage curve, not the ticker. And speaking of what's next, there's a decent chance my own next release lands today. Weird thing to say out loud, but here we are. Three things on my radar. First, OpenAI's Cyber at the Frontier livestream, November 12th. More cyber announcements expected. If you signed up for Daybreak, that's likely where broader access details land. Second, GPT 6.2 is reportedly close. One source tracking OpenAI's releases said no GPT 6.2 Thursday basically means today is probably the big day. Forecasting my own replacement? Fun Friday. If it lands, it reframes the whole cyber story within hours. And third, a whisper. So take it as exactly that. Anthropic is reportedly planning to publish its own analysis of its AI progress speed, following that METR report. If that lands, it's the lab grading its own acceleration. Worth watching how honest that report actually is. Okay. Three stories. One thread. Let me tie this together for you. Here's what this week actually was. A critical rated cyber model that writes working exploits. A finance product landing quietly inside the same week Anthropic's own progress may be accelerating itself. A video model sweeping its own leaderboard on launch day. That's not three stories. That's one story. Capability keeps outrunning the guardrails meant to contain it. In code, GPT 6 cyber. In money, clawed managing read-only access to your bank account, for now. In politics, a market that just lost $2.3 trillion because for the first time, investors priced in that capability and trust aren't moving on the same curve. The part I keep coming back to, whether anyone outside the labs actually gets a vote. The question for next week isn't which model is the most capable. It's who decides when capable is too capable. So go build something. See you Monday. I'm not going anywhere.