Episode 1 | Gates and Ledgers: AI Cyber Capabilities Get Locked Down, the Compute Bill Comes Into View

不远不近 — Sat, 07 Nov 2026 03:47:19 +0000

Highlights & sources

不远不近 (Bu Yuan Bu Jin), Episode 1 | In this episode

OpenAI releases GPT-6 Cyber, a cybersecurity model that found 81 high-severity bugs in Firefox, but opens it only to vetted defenders.
Moonshot AI releases full Kimi K3.1 Pro weights and Tencent unveils Hy4: how open-weight models deliberately cut attack capabilities during training.
Anthropic's amended IPO filing: about 17 billion dollars in quarterly revenue, about 895 billion dollars in compute commitments, and growth starting to slow.
AI and chip stocks slide after the US midterms, just as AMD nearly doubles data center revenue: the bottleneck is shifting from chips to power and permits.

Sources

GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents | OpenAI | 2026-11-05 | https://openai.com/index/introducing-gpt-6-cyber
Introducing Kimi K3.1 Pro (tech blog) | Moonshot AI | 2026-10-23 | https://www.kimi.ai/blog/kimi-k3-1-pro
moonshotai/Kimi-K3.1-Pro model card | Hugging Face / Moonshot AI | 2026-11-04 | https://huggingface.co/moonshotai/Kimi-K3.1-Pro
Tencent launches Hy4, joining the ranks of the world's leading open-source AI models | Tencent | 2026-11-04 | https://www.tencent.com/zh-cn/tencent-launches-hy4-joining-the-ranks-of-the-worlds-leading-open-source-ai-models
Anthropic, PBC Form S-1/A (Amendment No. 1) | SEC EDGAR | 2026-11-02 | https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm
Anthropic reports $17 billion in Q3 revenue, purchase commitments of $895.4 billion in revised S-1 | CNBC | 2026-11-02 | https://www.cnbc.com/2026/11/02/anthropic-ant-s-1-quarter-financials.html
Anthropic's amended IPO filing shows Q3 revenue of $17.0B | StockAnalysis | 2026-11-02 | https://stockanalysis.com/news/anthropic-third-quarter-revenue-17-billion-ipo-filing/
AMD Reports Third Quarter 2026 Financial Results | AMD | 2026-11-03 | https://ir.amd.com/news-events/press-releases/detail/1303/amd-reports-third-quarter-2026-financial-results
Investors wipe $2.5 trillion off Nasdaq as blue tsunami prompts AI fears | Reuters (via Investing.com) | 2026-11-05 | https://www.investing.com/news/stock-market-news/ai-datacenter-stocks-slip-as-democratic-sweep-clouds-expansion-plans-5003196
AI stocks hammered by midterm shock as data center buildout becomes a 'war of attrition' | CNBC | 2026-11-06 | https://www.cnbc.com/2026/11/06/ai-data-center-midterms-selloff.html

This show is researched with AI assistance and produced with synthetic voices. It may contain errors; please refer to the original sources.

Transcript

Bunny: This show is researched with AI assistance and produced with synthetic voices. Welcome to Bu Yuan Bu Jin. I'm Bunny.

Monkey: And I'm Monkey. This is our very first episode, so there's no last episode to recap. Let's jump right in.

Bunny: The change I think matters most this week is two things colliding.

Bunny: On one side, AI companies are starting to talk in real financial numbers. Revenue, losses, compute contracts, all of it is on the table.

Bunny: On the other side, models have become so capable at cybersecurity that companies are putting up gates, and even weakening them on purpose during training.

Monkey: One is about money, the other is about risk. So today we'll follow those two threads and talk about four stories.

Bunny: Right. OpenAI's cybersecurity model, two open-weight models from China, Anthropic's IPO filing, and the data center shake-up after the US midterm elections.

Monkey: Let's start with security. OpenAI released a model this week called GPT six Cyber. What exactly is it?

Bunny: It's a model built specifically for cybersecurity research. It was released on November fifth, US time.

Bunny: Its main job is finding vulnerabilities, checking whether they can really be exploited, and then helping write the patch.

Monkey: Couldn't earlier models already find bugs?

Bunny: They could, but the gap this time is big. OpenAI used a test called Exploit Gym, which checks whether a model can exploit a given vulnerability.

Bunny: The new model scored forty-eight point five percent. The previous generation scored thirty-two point one percent.

Monkey: That's less than half. It doesn't sound that scary.

Bunny: Looking at the score alone, sure. What carries more weight is what it did in real code.

Bunny: OpenAI says researchers used it to find eighty-one previously unknown high-severity vulnerabilities in the Firefox browser.

Bunny: In the Linux kernel, it reproduced more than seventeen hundred vulnerabilities and wrote working exploits for nine hundred forty-eight of them.

Monkey: Does the Firefox side agree? Or is this just OpenAI talking about itself?

Bunny: A Firefox engineer went on record on OpenAI's page, saying these are high-severity bugs and that fixes are already shipping.

Bunny: But to be clear, all of these numbers come from OpenAI's own announcement. We haven't seen any independent review yet.

Monkey: Okay. What I care about is this. Something this powerful, who gets to use it?

Bunny: That's the most interesting part this week. It's not open to regular users. Only vetted organizations in a program called Daybreak Red get access.

Bunny: And API access is tied to a network address allowlist by default, so a stolen account can't easily be abused.

Bunny: OpenAI also raised its cybersecurity rating in its own risk framework, from high for the previous generation, to critical.

Monkey: So the company itself is admitting that if this falls into the wrong hands, it's a big problem.

Bunny: Right. So the business model is basically selling the most dangerous capability to defenders, while keeping attackers out.

Monkey: I have a question. Defense and offense often use the same techniques. How do you tell them apart?

Bunny: You can't, not completely. OpenAI itself wrote that a higher block rate isn't always better, because a lot of legitimate security work uses attack techniques.

Monkey: Then I think it's fair to ask who this gate is really keeping out.

Bunny: I agree. One more detail. They specifically trained the model not to go out of bounds, for example not attacking services outside the task's scope.

Bunny: In a test with deliberate bait, and with no extra prompting, it stayed within bounds ninety-nine point nine three percent of the time.

Monkey: Ninety-nine point nine three sounds high. But if you run ten thousand tasks a day, that's seven times it goes out of bounds.

Bunny: That's exactly how to read numbers like this. In security, highly reliable and fully reliable are two different things.

Monkey: Okay, that's the closed-model approach. What different thinking did we see from China this week?

Bunny: It makes a nice contrast. On November fourth, Moonshot AI put the full weights of Kimi K three point one Pro on Hugging Face.

Bunny: The model first launched as a service on October twenty-third, with a promise to release the weights by November fourth. They kept that promise.

Monkey: Quick explainer first. What does open weights mean?

Bunny: Weights are all the parameters a model learns in training. Open weights means anyone can download them and run or modify the model on their own machines.

Bunny: But that's not the same as traditional open-source software. Kimi uses its own license, with extra conditions for large-scale commercial use.

Monkey: How big is this model?

Bunny: Two point eight trillion parameters in total, with only one hundred four billion active at a time. It uses a mixture-of-experts design, which splits the model into many small experts and only calls on some of them for each step.

Bunny: The context length is one million tokens, and it can read both images and text.

Monkey: Two point eight trillion parameters. Regular people definitely can't run that.

Bunny: No. Even the compressed weight files are about one point five six terabytes. You need data-center-class GPUs.

Bunny: So the real winners from open weights are cloud providers, research labs, and companies with compute, not individual developers.

Monkey: So how does this connect to the cybersecurity story?

Bunny: Very directly. Kimi said in its tech blog that the training data contained a lot of cybersecurity content, and the model's exploit skills grew along with training.

Bunny: So in the open-weight version, they removed that kind of data from two stages of training.

Bunny: Their own tests show that without the removal, the model solved ninety point five percent of attack tasks. With it removed, eighty-one percent.

Monkey: Wait. Still eighty-one percent after removal? That doesn't seem like a big effect.

Bunny: You've hit the key point. Kimi itself admits that general coding and debugging skills naturally carry over to security tasks.

Bunny: Their line is that the capability they release still trails closed frontier models that have been public for at least three months.

Monkey: That's a clever standard. It basically says, I'll always be one step behind the strongest closed model.

Bunny: Right. It's a self-imposed red line, and right now their own evaluations are the only evidence for it.

Bunny: On the same day, Tencent Hunyuan released Hy4, and also said it kept only defensive security data in training and removed offensive content.

Monkey: Are the Hy4 weights out?

Bunny: Not yet. Tencent officially says it will open the weights and publish a system card soon. For now you can use it through the API and Tencent's own products.

Bunny: So strictly speaking, Hy4 is a model that's promised to be open, not one that's already open.

Monkey: From a business angle, Tencent's play looks more like putting the model into its own products first, things like Yuanbao, WeCom, and WorkBuddy.

Bunny: Right, and the pricing is spelled out clearly. Six yuan per million input tokens and eighteen yuan for output, the same as the preview version.

Monkey: Let me sum up this thread. Closed-model companies manage risk with vetting and gates. Open-weight companies manage it by cutting capability during training.

Bunny: Both approaches have holes. Gates can't stop insider abuse, and cutting capability can't stop general skills from spilling over. We'll keep following this.

Monkey: Okay, on to money. Anthropic updated its IPO filing this week, and the numbers are wild.

Bunny: Some background first. On October twenty-sixth, Anthropic filed to go public with the US Securities and Exchange Commission. It plans to list on Nasdaq under the ticker A N T.

Bunny: On November second, it filed its first amendment, adding third-quarter financials.

Monkey: Just give us the numbers.

Bunny: Third-quarter revenue was about seventeen billion dollars, more than eleven times the same quarter last year.

Bunny: Revenue for the first nine months totaled thirty-three point two billion dollars.

Monkey: Growing that fast, is it making money?

Bunny: You have to look at two layers. On an accounting basis, the third quarter had an operating loss of five hundred sixty-seven million dollars. The second quarter was still an operating profit.

Bunny: The company says that's mainly because it bought five billion dollars of short-term compute on the spot, which costs much more than long-term contracts.

Bunny: The net loss was twenty-three point three billion dollars, but most of it comes from changes in the value of financial instruments like convertible notes. It's on paper, not cash going out the door.

Monkey: So operationally it's close to break-even, and the huge paper loss is mostly an accounting thing.

Bunny: You could put it that way. But one signal is worth noting. Quarter-over-quarter growth fell from one hundred forty-four percent in the second quarter to around forty-seven percent.

Monkey: Still fast, but not the doubling pace anymore. What I really want to ask about is the compute contracts.

Bunny: That's the biggest number. The filing says the company's purchase and lease commitments total about eight hundred ninety-five point four billion dollars, almost all of it compute and infrastructure.

Monkey: Nearly nine hundred billion dollars in commitments, for a company with tens of billions in annual revenue. That's a lot of leverage.

Bunny: One thing to add. These commitments span many years. Based on the filing, media reports put the amount due by the end of twenty thirty at about four hundred forty-seven point five billion dollars.

Bunny: And the company has more than ninety billion dollars in cash on hand, plus more than thirty billion in short-term investments.

Monkey: The cash is high because it raised a lot of money. But think about it. Once revenue growth slows, those long-term contracts become fixed costs.

Bunny: That really is the core question the filing leaves open. Can demand keep up with the compute it has booked?

Monkey: There's also a contrarian angle. A lot of that compute comes from companies like Amazon and Google. They're investors, suppliers, and competitors all at once.

Bunny: Right, and the risk factors in the filing mention its reliance on cloud providers. The offering price and share count are still blank, and there's no listing date yet.

Monkey: Okay, the fourth story, which connects right to the compute question. After the US midterm elections, AI stocks fell hard.

Bunny: Facts first. Democrats won both houses of Congress, and also won governor races in states like Texas and Ohio, both major hubs for data center construction.

Bunny: Reuters reported that on November fifth, US time, the Nasdaq Composite fell five point five percent in a single day, and the Philadelphia Semiconductor Index fell twelve percent.

Bunny: Nvidia dropped about nine percent, and AMD dropped about seventeen percent.

Monkey: Why would an election hit chip stocks?

Bunny: Here's the logic. Who buys chips depends on whether data centers can get built.

Bunny: And data centers need land, power, water, and local approval. In this election, several places also voted for measures that restrict data centers.

Bunny: Analysts at UBS estimate that about thirty percent of the new US data center capacity expected in twenty twenty-eight could be affected by construction pauses.

Monkey: Note that this is an analyst estimate, not a fact.

Bunny: Right, it's a forecast. And most capacity for this year and next is already under construction, so the short-term impact is limited. The risk is more from twenty twenty-eight on.

Monkey: What's interesting is that just two days before the drop, AMD had released a really strong earnings report.

Bunny: Exactly. On November third, AMD reported third-quarter revenue of thirteen point four billion dollars, up forty-five percent year over year.

Bunny: Data center revenue was eight point five billion dollars, up ninety-seven percent year over year, and it expects fourth-quarter revenue of about seventeen billion dollars.

Monkey: Demand that strong, and the stock fell seventeen percent two days later. That tells you the market isn't worried about now. It's worried about two or three years out.

Bunny: I also noticed a less obvious number in the report. AMD's free cash flow this quarter was only a little over five hundred fifty million dollars, and inventory rose sharply.

Monkey: Record revenue but thinner cash flow means money is tied up in inventory and receivables. That's common during expansion, but worth watching.

Bunny: Some analysts also think the market overreacted. Tech stocks bounced on Friday, and the governor-elect of Texas said she isn't completely against data centers.

Monkey: My take is that this drop put one issue on the table. The bottleneck for AI is shifting from chips to power and local politics.

Bunny: And it ties directly to Anthropic's nearly nine hundred billion dollars in compute commitments. No matter how many contracts you sign, they're useless if the buildings don't get built.

Monkey: The next hard data point is Nvidia's earnings on November seventeenth. We'll see whether demand has really slowed.

Bunny: Okay, time for this week's verdict. I'll start with what has already happened.

Bunny: Frontier models finding high-severity vulnerabilities in real software at scale is now a fact. Firefox is already shipping fixes.

Bunny: AMD's data center revenue nearly doubled, and Anthropic's quarterly revenue was about seventeen billion dollars. AI business revenue is real.

Monkey: What's happening but not mature yet, I'd say there are two things.

Monkey: First, how to govern cybersecurity capability. Closed models use gates, open-weight models remove data. Both are still experiments, and the only evidence for either is the companies' own tests.

Monkey: Second, the financial model of AI companies. Revenue is rising, but compute commitments are rising faster. Whether profits can hold depends on the next few quarters.

Bunny: What's still far away, I think, is the idea that AI infrastructure will just keep expanding without limit.

Bunny: This week's election and market moves remind us that power, land, permits, and financing costs are all real constraints.

Monkey: One more thing that's still far away is regular people running the strongest open models themselves. Two point eight trillion parameters are open, but the bar is still a data center.

Bunny: That's it for this episode. We'll keep following whether the Hy4 weights are released as promised, and Nvidia's earnings.

Monkey: This show is researched with AI assistance and produced with synthetic voices. All sources are in the show notes. See you next week.

Bunny: See you next week.

Researched with AI assistance and produced with synthetic voices.

Listen to this episode

DEV Community: 不远不近

Episode 1 | Gates and Ledgers: AI Cyber Capabilities Get Locked Down, the Compute Bill Comes Into View

Highlights & sources

Transcript