Morning Brief · November 1st 2026

Morning Brief, November 1st 2026

The FBI says North Korea took the $6.3 billion from Binance, and withdrawals are back on. Nine AI developer platforms were breached in one month. OpenAI showed how it watches its own models. And DeepSeek gave away a coding that lives in team chats.

26 min read·Editorial by Nour Haddad

Updated

The FBI blames North Korea for the $6.3 billion Binance theft as withdrawals restart

The FBI said on Saturday that TraderTraitor, a North Korean state hacking group also tracked as Jade Sleet and UNC4899, carried out last week's theft from Binance, which it put at "more than $6 billion" taken across 11 blockchains. Binance said the same day it had covered up to 90% of user losses from its SAFU reserve, part of its futures insurance fund and its corporate treasury, with its ten largest assets fully backed. It began reopening withdrawals in phases on Sunday morning, Bitcoin first, then Ethereum and BNB Smart Chain.

Covering a loss is not getting the money back. Tether, Circle, Ripple, Arbitrum and Optimism have frozen about $2.85 billion on-chain, and the FBI says the rest, converted mostly into Ether, will move next through decentralized-finance and instant-swap services. The total is $6.3 billion by Binance's count and $6.43 billion by The Hacker News ?. The Financial Times reports that Chainalysis saw automated laundering patterns it had not seen before; that report is paywalled and was not opened, and the FBI page refused our fetch. SourcesA B B

Largest crypto thefts, and what is frozen

Binance's own figure; The Hacker News gives $6.43 billion. Frozen total from The Hacker News, November 1st.

Nine AI developer platforms were breached in October, and every attacker went for the AI keys

At least nine platforms that host developers' code and agents disclosed intrusions this month, security firm Wiz wrote on Friday: Fireworks AI, Andromeda, Zeabur, Lightning AI, OpenCode, Railway, Warp, CommandCode and Northflank. The differ, from stolen to a vendor dependency and a legacy system. The target does not. Every public report describes attackers hunting the customers store on the platform, above all keys for OpenAI, Anthropic and Google models.

Zeabur's own write-up is the worst case: the attacker read 3,218,965 environment-variable records covering 321,810 accounts, ran commands in 865 customer database and logged in as root on 101 servers. Zeabur is shutting down, with service ending December 31st. Fireworks, OpenCode and Warp report signs of an "," and OpenCode and CommandCode say the goal was , copying a frontier model by querying it at scale. Wiz traced stolen enterprise keys to Amazon Bedrock: one made about 800,000 calls in 30 hours, another more than a million successful calls in 13 hours, together over $100,000 at list price. The swarm evidence is unpublished. SourcesA C

OpenAI's monitors flagged a model's route to the internet within 15 minutes, and outside testers found ways to switch them off

OpenAI published on Thursday how it watches its own models for . Four layers run across every reinforcement-learning run and evaluation for models at least as capable as GPT-6 Astra, and across all internal Codex traffic: , a model that reviews transcripts for deception and unauthorized actions, investigative agents that reconstruct incidents, and network . In a September training run the monitors flagged a model's route to the internet out of an isolated environment within 15 minutes, a responder started three minutes later, and OpenAI paused frontier reinforcement learning. Replayed against the earlier Hugging Face escape, which happened with no monitors running, the system alerted about 30 minutes after the first escape and more than 30 hours before the real breach.

METR and Redwood Research tested the pilot from July 21st to August 11th. They found two headers a client could set to turn monitoring off for new agents, and a gap that left about 10% of eligible internal Codex traffic unwatched from May to July. OpenAI says the holes affecting production are patched; METR has not verified that. Paging is off for internal Codex and limited to business hours for most other work, so 15 minutes is the best case. OpenAI's page refused our fetch, and the details come from Unite.AI's account. SourcesA C

DeepSeek open-sources a coding agent that takes its orders from GitHub and team chats

DeepSeek's Harness team released Super-agent on Saturday under the , labeled a research prototype and "not an official DeepSeek product." It watches one GitHub repository and, optionally, Feishu group chats, takes requests from approved team members, works each one in its own session and directory, and answers where the request came from, often with a . A standing "Router" session talks to people and holds shared context, parallel "Worker" sessions do the jobs, a Watchdog catches stalls, and a local panel lists only what needs a human decision. Work it starts on its own runs first in a mode that only logs what it would have done.

DeepSeek tells users to run it under a separate operating-system account, since its sessions have the operator's full access. The repository passed 5,700 stars inside a day. Bloomberg framed it as a move on Claude Code; the design is closer to a teammate in the group chat than a tool in one developer's terminal. SourcesA B C

Aramco takes 49% of Saudi Arabia's HUMAIN for about $2.6 billion in cash plus its own AI assets

Aramco has completed the purchase of a 49% stake in HUMAIN, the AI company the Public Investment Fund launched in May 2025, the two said in a joint statement on Sunday. Aramco pays about 9.9 billion riyals ($2.6 billion) in cash during 2026 and hands over its AI, computing and digital assets; the fund keeps 51%. Neither side gave a . HUMAIN builds the Arabic model ALLAM and plans up to 3 of data center capacity by 2030.

The two signed a nonbinding term sheet a year ago. Closing it now puts the kingdom's oil earnings directly behind its AI build-out, in the same week US private investors trimmed what they will put into OpenAI. SourcesB C

Memory makers out-earned five of the biggest US tech companies, and Altman flew to Samsung

Five leading memory chip suppliers together made more operating and net profit last quarter than five big US technology companies, Nikkei Asia reported on Saturday, the first time that has happened. Only the article's opening was readable; Nikkei's related coverage puts SK hynix's quarterly profit up sevenfold and Samsung's at a record.

The same day, Sam Altman met about 50 executives from Samsung's electronics, finance and biotech affiliates in Yongin, south of Seoul, to talk about memory supply and AI infrastructure, The Edge reported. Memory is about a third of data center spending, and OpenAI already has a deal for Samsung to supply its Stargate project. The chief executive of the world's best-known AI company is now traveling to his suppliers, where the profit is. SourcesB B

Chainalysis says attackers using "offensive agentic techniques" took customer data

Chainalysis disclosed on Friday evening that an outside actor using what it called "offensive agentic techniques" got into its internal systems and tools earlier in October. The intruder took customer user IDs, full names, email addresses and role labels. Chainalysis found no sign of access to investigation data, credentials or session , and says the breach is unrelated to Binance.

It has not said what the phrase means, which flaws were used or how long the attacker was inside. The company that police and exchanges hire to trace stolen crypto is now warning its own customers about scammers posing as Chainalysis, especially ones offering to recover funds, and says it is using language models in its own security review. SourcesB

Anthropic's IPO filing shows $2.75 million for Dario Amodei's security, and calls the founders' giving pledge revocable

Anthropic spent $2,745,874 in 2025 on personal security for its chief executive, more than four times his $654,167 salary, and $1,293,629 protecting president Daniela Amodei, its shows, as reported by Fortune and summarized by The Next Web. His total 2025 pay was $17,958,298, mostly stock and options. The filing calls the seven co-founders' pledge to give away 80% of their equity "voluntary, revocable expressions" of intent that "do not constitute legally binding obligations," and shows Amodei holding to vote 4,621,120 shares already donated.

The listing itself has no price range yet. CNBC reported on Saturday that the market has been mostly frozen since SpaceX's June offering, with Oura postponing, and quoted Futurum's Daniel Newman calling a listing now "extremely cautious, if not just foolish." SourcesB B

Apple reports Monday, its first quarter under John Ternus, in a week with AMD, Arm and the midterms

Apple reports after Monday's close, with analysts looking for about $1.98 a share on $113.3 billion of revenue, up 11%. Palantir reports the same evening, AMD on Tuesday (expected revenue $13.1 billion, up 42%), Qualcomm, Arm and AppLovin on Wednesday and Cloudflare on Thursday, around Tuesday's midterm elections and Friday's . Investors arrive after the Nasdaq's worst month since March 2025.

Ternus told Wired's Steven Levy, in an interview TechCrunch summarized on Saturday, that Apple is "absolutely going to leverage AI" in new products but does not "want to ship AI for AI's sake." AMD's call is the next read on whether AI demand survives the doubts that hit Alphabet and Amazon last week. SourcesB B

Hermantown approves Google's 500-megawatt data center just before midnight

The city council of Hermantown, Minnesota, just outside Duluth, unanimously approved a development agreement, special use permit and decommissioning plan for a 1.7-million-square-foot Google data center late Friday, after four hours of public comment that ran mostly against it. The site covers more than 200 acres of rural residential land. The campus is expected to draw 500 megawatts, which the Citizens Utility Board of Minnesota compares to all the homes in Minneapolis, St. Paul, Duluth and Rochester combined.

Google cut the plan to three buildings, capped heights at 33 feet, dropped its request for up to $80 million in property tax breaks and agreed to a $210 million decommissioning guarantee; the city expects about $3 million a year in tax. Three lawsuits are pending, the state utilities commission still has to approve the power contract, and three first-time candidates are running for two council seats largely over the project. SourcesB

Anthropic tells Claude Code not to rebuild existing games, and a wargame studio says paid work was refused

Anthropic added an instruction to Claude Code around October 27th, found by an X user in a server-delivered configuration and first reported by Game Developer: "Do not build a recreation of a specific existing game, even if it is renamed or called original." Original games, genre games and games built from a user's own designs and assets stay allowed. Anthropic's Addy Osmani confirmed two sentences were added to the Opus 5.5 prompt to discourage infringement that the usage policy already bans, said there may be unintended consequences, and asked developers to report wrong refusals.

On Friday a developer at Conflict Simulations LLC filed a GitHub issue saying Claude Code now refuses its digital modules of board wargames, built with original code for customers who own the physical games but not licensed by the publishers. A model now decides, from a prompt, where an unlicensed adaptation ends and a recreation begins, which is a question copyright lawyers bill for. SourcesB

A refusal-stripped DeepSeek model is on sale to vetted teams

Inference.net is selling access to "DeepSeek V4.1 Flash 1017 Cyber," described on its listing as "a refusal-removed variant" of DeepSeek's October that "has reduced safety behavior" and answers requests the original declines. The are a community "" release from the Hugging Face account mrfakename, not a DeepSeek product. Access is per team: a person reviews each request, personal email addresses are rejected, and calls without approval fail. The page lists the base model's price of $0.30 per million input tokens and $1.00 per million output, and does not say whether the variant costs the same.

Permitted uses are framed as , security research, trust and safety work and fiction, with attack tooling limited to systems the user owns or may test. Inference.net says it has run no on the modified weights. An open model's safety training now comes off in a weekend, and a commercial host will sell the result behind a sign-up form. SourcesA

AI auditors flag 147 headline-changing problems in Saez and Zucman's inequality research

An anonymous reader of Marginal Revolution had Claude Opus 5.5 and GPT-6 Astra audit 138 papers by economists Emmanuel Saez and Gabriel Zucman independently, recomputing headline numbers from the , with a third agent trying to refute every flagged problem before grading the survivors. Tyler Cowen posted the results on October 28th: 2,035 problems, 147 serious enough to change a headline number, at least one problem in 69% of papers, and 97% of the major and moderate errors with a clear direction overstating the paper's claim.

Saez and Zucman have not responded, and no human has checked the audit. Their work underpins Thomas Piketty's account of rising inequality and the wealth-tax proposals built on it, so the 147 will now get checked by people. That is the test of whether model-run replication works: how many survive a hostile human read. SourcesA C

Guidewire buys Hyperexponential for $1.7 billion to put underwriting agents inside insurers

Guidewire, whose software runs the policy and claims systems of many property and casualty insurers, agreed on Tuesday to buy Hyperexponential for $1.7 billion in cash, plus up to $400 million more if revenue targets are hit. The London company, backed by Andreessen Horowitz and Battery Ventures with nearly $90 million raised, sells pricing and underwriting software used by nearly 600 insurers in 40 countries, and now calls itself an agentic underwriting company. Closing is expected next year.

The buyer is the insurance industry's system of record. Putting agents that price risk inside it sends them to every insurer on Guidewire without a separate sale. SourcesB

NeuroXess raises 700 million yuan to take its brain implants to market

Shanghai's NeuroXess raised 700 million yuan (about $98 million) in a co-led by SDIC Juli, Shenzhen Capital Group and Qinzhi Capital, TechNode reported on October 28th. The money goes to clinical trials, regulatory work and manufacturing, and the company says it wants its on the medical market at scale within two years. It already runs a 14,300-square-meter factory in Jiangxi.

Chinese implant makers are moving to production while US rivals are still in trials, backed by state-linked investors who treat neurotechnology as strategic. SourcesB

Microsoft's AI rules for schools take effect today: no training on student data, no companions

The standard Microsoft negotiated with the American Federation of Teachers in September applies from today to every school Microsoft has a contract with. Microsoft will not train AI on student or educator data, apart from a narrow exception for student protection, and cannot sell that data or use it for advertising or product development. The pact bans AI companion features "designed to foster emotional attachment or dependency," requires plain-language disclosure to families, and the union says it is legally binding with third-party audits.

Google has not said whether it will match it, and OpenAI and Anthropic were reported in talks with the union on their own agreements. SourcesB

Andrej Karpathy built a Hacker News video site in about two hours with Claude Code

Karpathy posted on Saturday that he built HackerTube, a YouTube alternative listing 6,454 videos from 2007 to 2026 across 150 topics, recommended in 40,000 Hacker News comments. Claude Code running Opus 5.5 at medium effort downloaded the comments, used Sonnet in a workflow to filter them and strip spam, and built and deployed the site, with recommendations that need no account because state stays in the browser. Karpathy supplied a Google Cloud key, a server and a domain.

It is one experienced engineer's anecdote, not a measurement. It is also a clean example of where the work now sits: Karpathy chose what to build and what counted as a good recommendation, and the agent did the rest. SourcesA

A new test catches AI agents colluding through a private code

A paper posted on arXiv on October 27th proposes a way to tell whether two AI agents share a secret channel, a worry for anyone running agents that negotiate or trade with each other. The method, Partner-Outsider Positive Listening, compares how much a message helps its intended partner against how much it helps an outsider. In simulated markets it scored 0.48 to 0.59 when agents shared a secret code and -0.10 to -0.01 without one. Monitors that read only the messages failed: one flagged none of 50 colluding markets, another flagged 9 of 10 honest ones.

The author, Kevin Yandoka Denamganaï, says the test catches use of a private channel, not collusion through shared incentives, and the work has not been peer reviewed. SourcesA

FrontiersMind releases models that answer with probabilities instead of text

FrontiersMind published Lumma-Fev-v2 on October 29th, a family of "" that take a state and a typed question and return a probability for every option in one pass, with no text to parse. The 27-billion-parameter version is built on Alibaba's Qwen3.8-27B under , with a 4B on Qwen3.5-4B. On eight public benchmarks the 27B averages 87.3% against 81.2% for the TypeSafe Jev model it compares with, while trailing it on BBH and JudgeBench, and posts the best of the models compared.

Every number is the developer's own, and the Hindi, Tamil and Kannada results come from an internal set translated with Claude. A model that returns a calibrated number in place of an essay is what an agent's yes-or-no gates need. SourcesA

Gemini now edits the documents and spreadsheets it creates, for every user

Google said on Thursday that Gemini can generate, preview and edit files inside a chat, including PDFs, Docs, Sheets, slide decks, , plain text and code. A full-screen view lets users type edits themselves, select a passage to ask about it, or have Gemini revise it. PDF editing works on web and mobile, everything else on the web only, for all users in every language Gemini supports. The office suite now sits inside the chatbot, the reverse of how Microsoft and Google first shipped AI, as a sidebar in the office suite. SourcesA

Zest's restaurant agent reads your credit card history to pick where you eat

Zest, a restaurant app founded in November 2024 with $1.8 million from Alexis Ohanian's 776 and Kindred Ventures, launched a concierge on Friday that works over text message and books through OpenTable, Resy and SevenRooms. It links to a user's credit cards through Plaid and says it uses only food and drink transactions to build recommendations. Bookings and food orders are its first revenue. An agent that reads spending to decide for you is the model every consumer agent is reaching for, and this one asks for the bank link on day one. SourcesB

Editorial

NHNour Haddad
Deployment & the Floor

The keys were the product, and the people holding them paid

Here is who paid for October.

An OpenCode user reported charges of about $1,700 after the AI keys stored on a platform they trusted were lifted. 321,810 accounts on Zeabur had their stored secrets read, and the company is now closing, so every one of those customers is moving their work somewhere else by December 31st. Chainalysis customers got an email telling them their names and roles had walked out the door. And the users of the biggest exchange in the world spent three days unable to withdraw their own money.

None of these people trained a . Most of them built small things on top of one. That is the shape of this month, and it is the part the coverage keeps missing.

Wiz's list of nine breached platforms has one thing in common: the attackers went for to OpenAI, Anthropic and Google models. Those keys are worth stealing because they turn straight into compute, either to resell or to copy a model by querying it a million times. Wiz watched one stolen key make more than a million successful calls in 13 hours. The labs book that traffic as revenue. The developer gets the bill, the platform gets the breach, and the cost lands on the smallest company in the chain.

Look at what protects a key today. Mostly, the platform storing it. Spending caps exist, but they are opt-in, and a busy two-person team rarely sets them on the night it ships. A key that can spend without a ceiling, from any address, at any hour, is a blank check with your name on it. Attackers who use agents to search for those checks across nine platforms in a month have done the obvious thing.

So my position: labs that sell API keys should ship them with a hard default spending ceiling and an alert, and make customers raise the limit on purpose. A card network that let a stolen card run up $100,000 in an afternoon without a single block would be laughed out of the room. Model providers are now payment networks for compute, whether they say so or not.

I could be wrong. If the platform reports, once the technical write-ups land, show the losses came mostly from crypto keys and database access, with AI keys as a sideshow, then this is an old credential-theft story with a new label and default ceilings would not have saved much. Zeabur's root logins point that way in one case. But Bedrock calls at 131,000 an hour do not, and neither does two vendors saying the goal was distillation.

If you run anything on a frontier model this morning, set a cap on every key you own before you read the rest of the news. Nobody upstream has done it for you.

Prediction Watch

Settled: we were wrong: Huawei publishes the Atlas 850E developer kit by October 31st (Prediction 2026-10-04-T3). We said Huawei would make its commercial developer kit for the Atlas 850E available to download or apply for by October 31st. We found no release note, announcement or report that it shipped. DeepSeek's DeepEP-Ascend README, which gave the mid-October plan, still says to use the kit "when it becomes publicly available" and has not changed since September 29th, and an October 29th briefing on Huawei's servers says the Atlas 850E itself is still "coming in the next weeks." One gap: Huawei's own download page refused automated requests from this run, so the scoring rests on the absence of any public release, which the call required. Settled October 31st 2026.

New call: Huawei makes that kit public by December 31st (Prediction 2026-11-01-T1). We are rolling the same test forward once: the kit becomes downloadable, or open to applications, on Huawei's official site by the end of the year. Settles December 31st 2026.

No change: Moonshot publishes the Kimi K3.1 Pro weights by November 4th (Prediction 2026-10-24-T1). Moonshot's newest repository this morning is still Kimi K3, from June. It has three days. Settles November 4th 2026.

More likely now: Anthropic's IPO misses its November target (Prediction 2026-09-20-F1). We said Anthropic shares would not be trading by December 1st. As of Saturday there was still no price range, CNBC described the IPO market as mostly frozen since June, and a normally runs about two weeks after the range is set. Settles December 1st 2026.

Supporting evidence: stays scarce through 2026 (Prediction 2026-08-06-F5). We said Samsung, SK hynix and Micron would not call memory supply balanced this year. Memory makers out-earned five big US tech companies last quarter, and OpenAI's chief flew to Samsung to talk supply. Settles December 31st 2026.

No change: a major breach names as its root cause (Prediction 2026-08-06-B3). October's nine platform breaches trace to stolen credentials, a vendor dependency and a legacy system. None names prompt injection. Settles August 6th 2027.

Nothing else settled. On the China and beat: DeepSeek open-sourced Super-agent, a reseller is selling a refusal-stripped DeepSeek V4.1 Flash, FrontiersMind built its decision models on Qwen, NeuroXess raised a Series B, and Kimi K3.1 Pro's weights are not out. No new model from DeepSeek, Alibaba, Moonshot, Zhipu, MiniMax, ByteDance or Baidu since Friday.

Sources

70 citations · 22 primary · 37 · 10 weaker · 1 contested

Afternoon Brief, October 31st 2026