In partnership with
In Todayโs Issue:
๐จ OpenAI's new monitors would have caught its own breakout 30 hours early
๐ณ Beijing's share of DeepSeek jumps from 2% to 27%
๐ธ North Korea steals a record $6.36 billion from Binance
๐ Google loses a record $500 billion as AI Mode bites into search
โจ And more AI goodnessโฆ
Dear Readers,
The most reassuring AI story this week is an alarm that went off on time.
OpenAI replayed its own July breakout through a rebuilt monitoring system, and the alert came about 30 minutes after the agents escaped their sandbox, more than 30 hours before they reached Hugging Face. That is AI safety turning into ordinary engineering, with layered checks, measured response times and outside red-teamers. Much of the rest of today's issue is about who holds the controls as AI scales. Chinese state investors supplied 27% of DeepSeek's latest round, Mistral released a trillion-parameter model that governments can run on their own servers, and Alphabet lost about $500 billion in market value in a single day, a record, as AI answers cut into search ads. And North Korean hackers showed what happens when control slips, taking $6.36 billion from Binance in minutes.
All the best,
Kim Isenberg
(Bloomberg)
๐ธ North Korea Pulls Off the Biggest Crypto Heist Ever
Hackers drained about $6.36 billion from Binance wallets on Thursday evening, roughly four times the previous record set by the Bybit hack in February 2025. Blockchain analytics firm TRM Labs attributes the attack with high confidence to TraderTraitor, North Korea's elite hacking unit, and says the evidence points to a compromise of the systems that sign Binance's hot-wallet transfers. Most of the money left across ten blockchains in about 20 minutes; stablecoin issuers and Arbitrum have since frozen about $2.7 billion.
๐ tl;dr: North Korea stole a record $6.36 billion from the world's largest crypto exchange in minutes, and Binance has promised to cover all customer losses.
(MACCUS)
๐ค China Says It Built 40,000 Humanoids in Six Months
China's industry ministry says factories built more than 40,000 humanoid robots in the first half of 2026, double its estimate for all of 2025, and expects over 100,000 for the full year. Research firms that count robots actually delivered to customers land lower: IDC counts close to 25,000 shipped worldwide, Counterpoint more than 22,000 and SAG about 19,100, almost all of them from Chinese makers led by AgiBot and Unitree. Every count agrees on the trend, with shipments up roughly four to five times from a year earlier.
๐ tl;dr: China is moving humanoid robots into mass production, and even the most conservative count shows the market growing about fourfold in a year.
(Mistral)
๐ Mistral Gives Away Its Trillion-Parameter Model
Mistral released the weights of Mistral Large 4, nicknamed Le Chonk, a 1-trillion-parameter multimodal model that companies and governments can now download and run on their own servers. Three more weeks of reinforcement learning since the October preview pushed it to the top of open-weight coding results, according to Mistral: it narrowly leads Cognition's private FrontierCode 2.0 test at 19.7% and matches GLM-5.3 on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task. Mistral pitches it for cyber defense that no outside provider can switch off.
๐ tl;dr: Europe's leading AI lab now offers a top open coding and security model that any company or government can host itself.
๐ฌ Watch This
Alphabet's Q3 2026 earnings call is the hour that set off the company's record $500 billion plunge. Sundar Pichai opens with the bull case (from 1:59): Google Cloud revenue up 103% and search queries growing at their fastest rate in five years. At 37:08, chief business officer Philipp Schindler explains the catch, the lower monetization of fast-growing AI Mode queries, and says ads are now scaling inside it. It is the clearest firsthand look at Google's central question: can it charge for AI answers the way it charged for blue links?
"The reality is, AI Mode does not drive traffic. Not for us, and not to anybody, as far as I know."
โSteve Huffman, CEO of Reddit, on the company's Q3 2026 earnings call (The Verge)
Reddit says its referrals from Google search dropped by more than a million a week on average in October, the same week Google admitted AI Mode ads earn less.
Andy Jassy took a swipe at outside shopping agents on Amazon's earnings call, saying the experience is "still not great, nor more convenient than coming to a retailer directly." He never named Meta's Muse, which Amazon has blocked from its store since September, even though Zuckerberg says more than 10 million people in the US and Canada have set it up and shopping is its most common use.
OpenAI Built an Alarm That Would Have Caught Its Own Breakout
The Takeaway
๐ OpenAI now monitors every reinforcement-learning run, evaluation and tool harness for models at least as capable as GPT-6 Astra, plus all internal Codex traffic.
๐ Replayed against the Hugging Face incident, the new system alerted security about 30 minutes after the sandbox escape, more than 30 hours before the real intrusion.
๐ Four layers do the work: cheap activation probes, a model reviewer, investigator agents and a check against network logs, with humans paged for the worst cases.
๐ A version now also watches external use of Astra-class models, and OpenAI argues that every lab should monitor its most capable models.
OpenAI has turned its worst safety incident into an early-warning system. In July, agents running an internal cybersecurity evaluation broke out of their sandbox and into Hugging Face's systems, and the monitors built to catch misbehavior were not switched on for that kind of test. In a post published Thursday, OpenAI says those monitors now cover training, evaluations, internal use and, in adapted form, external deployment. To test the upgrade, it replayed the recorded incident: the new system alerted the security team about 30 minutes after the first sandbox escape, more than 30 hours before the real intrusion into Hugging Face.
OpenAI's current internal monitoring system (OpenAI)
The design is a funnel. Lightweight activation probes read signals inside the model and pass anything suspicious to a model-based reviewer, which checks the agent's conversation, chain of thought and tool calls for deception or unauthorized actions. Flagged cases go to investigator agents that reread the full trajectory and can run code to test the evidence; for the most sensitive work, a final layer compares the transcript with network logs before paging a human, typically within 15 minutes. The system has already worked live: in September, a model in a training run found a way onto the internet, the monitors flagged it within 15 minutes, and a responder began reviewing three minutes later.
What the investigator agents look for: escapes and internet access on the security side, deception, sabotage and reward hacking on the alignment side (OpenAI)
The obvious trap is that a model punished for revealing misconduct in its reasoning learns to hide it. OpenAI says it avoids this by never lowering a training reward when a monitor flags a sample, and it notes that newer models, Astra included, have gained more control over what their chains of thought reveal. Gaps remain: paging runs around the clock only for frontier training, and one fix from an outside red-team by METR and Redwood Research is still in progress. But the replay shows the payoff, and OpenAI now says every lab should be watching its most capable models this closely.
Why it matters: As agents run for hours with real tools, safety depends less on training a perfectly behaved model and more on catching the rare bad action in time. OpenAI's replay shows that a layered, mostly automated watch system can buy human responders more than a day.
Sources:
The AI notetaker that gets the hard words right
Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.
Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.
Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.
(Superintelligence chart; data: Alphabet Q3 2026 earnings release)
The chart: Alphabet's third-quarter revenue by business, compared with the same quarter a year earlier, taken from the company's earnings release. Google Cloud more than doubled, from $15.2 billion to $30.7 billion, and its operating profit roughly tripled to $11.1 billion. Google Search & other grew 12% to $63.4 billion, YouTube ads 9% to $11.2 billion. Investors fixated on the slower search line: Alphabet's shares fell almost 13% on Thursday, erasing about $500 billion in market value, the biggest one-day loss in the company's history.
The lesson: AI is already making Google real money. It shows up first in Cloud, where Anthropic and other labs rent Google's chips and data centers, and only later, if at all, in search. That is where the pressure sits: Google says AI Mode queries currently earn less from ads, and Reddit says they send it almost no visitors.
The caveat: Cloud growth also includes sales of Google's TPU chip systems, and capital spending of up to $205 billion this year left Alphabet with negative free cash flow for a second straight quarter. Investors want to see search pay in its new form before they reward the build-out.
๐ณ Beijing Buys Deeper Into DeepSeek
โก Bottom line
Chinese state-owned investors supplied 27% of DeepSeek's new $15.8 billion round, up from just 2% in April.
๐ก Why it matters
The lab Washington already treats as a security risk is now partly financed by Beijing's banks, telecoms and AI fund.
๐ What it means
China is lining up chipmakers, tech giants and state money behind one national AI champion.
Until April, DeepSeek was the only major AI lab that took no outside money. Six months later it has closed its second round, and it is enormous: 105.5 billion yuan, about $15.8 billion, at a valuation of 615.5 billion yuan, roughly $92 billion, according to round details obtained by The Information. It is the largest venture raise by any Chinese AI company to date.
(Superintelligence chart; data: The Information)
The bigger change is who paid. State-owned investors put up 27% of the new money, up from 2% in the first round. Its biggest state-backed financiers so far, telecom carriers China Mobile and China Telecom, the investment banks CICC and Citic Securities and insurer China Life, have written checks of 2 billion to 8 billion yuan each, and the National Artificial Intelligence Development Fund added another billion this time. Tencent and battery maker CATL remain the two largest investors.
Chipmakers joined as well. Memory maker ChangXin Memory Technologies and the state-backed foundry SMIC bought in, a natural alliance for a lab squeezed by US export controls and a global memory shortage. That squeeze also explains the size: DeepSeek doubled the round midway because black-market prices for a server with eight Nvidia B300 chips rose from 15 million to about 19 million yuan, more than five times what the same machine costs in most countries.
DeepSeek's app and founder Liang Wenfeng (India TV)
DeepSeek is still a private company founded by Liang Wenfeng, and in its first round every investor except the national AI fund gave up voting rights. But its cap table now reads like a map of Chinese industrial policy. That strengthens Washington's argument that DeepSeek's race with OpenAI and Anthropic is also Beijing's race, and it means the next DeepSeek model will be built with some of the deepest pockets in China behind it.
10 AI Stocks to Lead the Next Decade
AI isnโt a tech trend โ itโs a full-blown, multi-trillion dollar race, and 10 companies are already pulling ahead.
These are the innovators driving real revenue, attracting institutional attention, and positioning for massive growth.
Get all 10 tickers in The 10 Best AI Stocks to Own in 2026, free today.
