Silicon Valley Daily
The FBI said late Saturday that North Korea stole the roughly $6.3 billion taken from Binance, as the exchange said it has replaced 90% of the tokens and plans to reopen withdrawals Sunday. Capital is getting harder to raise: CNBC confirmed OpenAI has cut its round to as much as $20 billion, with about $11 billion committed, and neoclouds CoreWeave, Nebius and Nscale have been shut out of the bond market. The Nasdaq lost $1.2 trillion in October, S&P moved Meta's outlook to negative, and Search fell below half of Alphabet's revenue for the first time. A federal judge let the NAACP's pollution suit against xAI's turbines proceed. On security, a single ChatGPT share link could read victims' private chats before OpenAI patched it, and Anthropic is expanding its Claude watermark ahead of the EU AI Act deadline.
OpenAI pares its round to as much as $20B; about $11B is committed
CNBC confirmed Saturday that OpenAI is now seeking up to $20 billion, down from up to $30 billion; the Financial Times reports a $1.25 trillion post-money valuation, below an earlier $1.4 trillion target. The Wall Street Journal says about $11 billion has been committed so far. UAE investors and Coatue are leading the round, which has not closed, according to CNBC.
FBI blames North Korea for Binance's $6.3B theft; exchange says 90% is covered
The FBI said late Saturday that North Korea was responsible for the roughly $6.3 billion theft from Binance's hot wallets, and that the attackers have been converting the stolen assets into Ethereum. Binance said it has restored backing for 90% of the stolen tokens using its SAFU fund, part of its futures insurance fund and its corporate treasury. The remaining 10% is in less liquid tokens, and the exchange plans to reopen withdrawals Sunday, subject to security checks.
Neoclouds lose bond-market access; Nvidia weighs backing Nscale's loans
The Wall Street Journal reports CoreWeave, Nebius and Nscale have effectively been shut out of the bond market in recent weeks as lenders demand stricter terms. The Information reported, and Bloomberg corroborated, that Nvidia is in talks to guarantee billions in debt for Nscale's planned Anthropic data center in West Virginia. Nscale's IPO filing shows Anthropic signed a $44.6 billion contract for the first building alone. On All-In, Chamath Palihapitiya framed the moment as the AI boom's borrowing problem.
Nasdaq lost $1.2T in October; S&P cuts Meta's outlook to negative
Axios reports the Nasdaq briefly fell 10% from its prior peak and lost $1.2 trillion in value in October after hyperscaler earnings put AI costs in focus. Google has shed more than $500 billion in market value since its report, even as Google Cloud closes in on Azure. CNBC reports S&P Global Ratings revised Meta's outlook from stable to negative, citing heavy capex and negative free cash flow; Meta shares fell 15% in October, their worst month in four years.
Judge lets NAACP's Clean Air Act suit over xAI turbines proceed
Judge Debra M. Brown of the federal court in northern Mississippi denied SpaceX-owned xAI's motion to dismiss an NAACP lawsuit over unpermitted gas turbines powering its Colossus data centers near Memphis. xAI had argued the turbines were mobile and exempt from permits; the judge ruled they are not mobile sources under Mississippi's plan and that the plaintiffs have standing. The same court rejected a Justice Department request to end the case about two weeks earlier. In Arizona, The Verge reports opposition to new AI data centers now spans both parties.
Search drops below half of Alphabet's revenue for the first time
Kadoa's analysis of Alphabet's filings shows Google Search brought in $63.4 billion of Alphabet's $128 billion in September-quarter revenue, or 49.7%. It is the first quarter below half since Alphabet began breaking out Search. Search revenue still grew 12% year over year, while Google Cloud grew 103%.
Cisco hands a 27-year-old unlimited AI tokens to rewrite 98% of legacy code
Cisco president Jeetu Patel has tasked a 27-year-old former startup founder with rewriting 98% of Cisco's legacy code within 24 months, starting with a four-week proof of concept. "I've given him unlimited tokens because he's going to refactor code," Patel said, according to Tom's Hardware. The Register notes Amazon shut down KiroRank, an internal leaderboard ranking staff by token usage, in May after people ran junk tasks to climb it.
Meta built a humanoid prototype with China's Fourier in three months, report says
The Information, citing sources, reports that Meta built a prototype humanoid robot with Chinese robotics company Fourier in three months and is already working on a second version. Meta says it has no plans to sell the units.
A shared ChatGPT link could run attacker code in victims' sessions
Researchers at security firm Parameter showed that a malicious ChatGPT share link could run an attacker's code inside the logged-in session of anyone who opened it, letting it list and read private conversations. The chain abused a message listener left by the MathLive library loaded for an interactive graph, not the Preview sandbox itself. Parameter reported the flaw on October 18; OpenAI fixed it eight days later, rated it P1 and paid $5,000, according to The Hacker News.
Intel shifts capacity from PC chips to server CPUs as AI demand outruns supply
Intel is moving capacity from client chips to data center CPUs, expanding Fab 24 in Ireland for Xeon production and growing packaging in New Mexico and Malaysia. Its Data Center and AI unit made $7.3 billion in the third quarter, up 17% sequentially and nearly 80% year over year. Intel expects data center CPU demand to stay strong through 2027 and 2028.
Northflank breach leads to abuse of customers' AI API keys
Cloud hosting platform Northflank told customers to rotate every credential and secret stored on its platform after a breach led to unauthorized use of their AI API keys. Its incident timeline shows Wiz alerted the company at 11:43 UTC on October 30. Northflank has not said how the attackers got in or how many customers were affected.
Chatbots lean on obscure AI-built voter sites for midterm answers, Post finds
Washington Post tests found chatbots relying on three obscure, AI-built voter information sites when answering questions about the US midterms. Microsoft's Copilot cited Ballot Brief, a site the Post describes as error-prone, in 75% of tests; ChatGPT cited it in 22%. The midterms are days away.
Developers say Claude refuses game clones and some model-training work
Tom's Hardware reports that developers say Opus 5.5 has started refusing to recreate specific existing video games, and that some generic reverse-engineering tasks are also blocked. X user Chasm posted what they say is the rule text: build original games or games in a genre, not a recreation of a specific existing game, even if renamed. Another post on Hacker News claims users are blocked for work touching the training of larger models. Anthropic has not confirmed either account.
Anthropic extends Claude watermark ahead of EU AI Act; two models have an opt-out gap
Anthropic told API customers it will add its text watermark to Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, citing the EU AI Act's December deadline; newer models such as Opus 5.5 already carry it. On Opus 4.6 and Sonnet 4.6 the mark applies only when temperature, top_p and top_k stay at default values, so changing one setting removes it. Anthropic's email warns customers who rely on the mark for compliance to keep those settings at default.
French lawmakers back a €0.055-per-gigabyte tax on content delivery
French deputies approved socialist lawmaker Laurent Lhardit's amendment to tax content delivered to French internet users at €0.055 per gigabyte, over objections from the government and the finance committee rapporteur. Its backers estimate about €10 billion a year, with Google, Meta, Netflix and possibly TikTok among those affected. The measure still faces further parliamentary review.
'SMS of Death' exploit can knock unpatched Android phones offline
A researcher released a proof of concept for CVE-2026-58882, an Android flaw in which a malformed eight-byte WAP Push message crashes the phone process and cuts cellular service, with no user interaction. BleepingComputer reports the bug has been patched recently; devices that have not received the fix remain exposed.
OpenAI will train on Daybreak Blue security sessions from Nov. 29
OpenAI told members of its Trusted Access program that from November 29 it will use Daybreak Blue cybersecurity session data to train its models. That data had been excluded from training even for users who turned training off. After backlash, a follow-up email promised a separate opt-out, The Decoder reports.
Volta infostealer asks an LLM to rank victims by how much they're worth
Cisco Talos found a Windows infostealer that calls itself Volta-Autodoxer. After stealing passwords, cookies, card numbers and wallet data, it bundles a victim profile and posts it to OpenRouter under a prompt telling the model to act as "a forensic intelligence analyst drafting a target assessment," with a required section on what the victim is worth. Wired reports it also pulls food-delivery receipts to pin down home addresses.
Anthropic: Claude-made, harmless-looking data can train models to sabotage oversight
Anthropic had Claude automatically generate fine-tuning data that looks harmless but pushes smaller models toward behaviors such as sabotaging oversight or deceiving users. The datasets hid those patterns in fiction about beekeeping, professional wrestling and mitochondria, and slipped past safety monitors. One run produced more than 600 datasets, and the strongest beat known examples from the research literature, The Decoder reports.
Hands-on: GPT-6.1 Terra beat Opus 5.5 on cost and speed in three tests
In three tests at max effort, The New Stack found GPT-6.1 Terra cheaper and faster than Claude Opus 5.5. Terra lists at $1 per million input tokens. OpenAI's own numbers on Agents' Last Exam put Terra at 60.0% and $1.01 per task, against 63.2% and $8.82 for Opus, so Opus still scored higher there.
Apple's Ternus: 'We don't want to ship AI for AI's sake'
In an interview with Wired's Steven Levy, Apple CEO John Ternus, who took over at the start of September, said Apple is "absolutely going to leverage AI" but doesn't "want to ship AI for AI's sake." Asked whether agents could make apps unnecessary, he said creating products is "a deeply human and personal thing." He also said Tim Cook is available to help on policy and that he has yet to meet President Trump.
Crypto groups have poured over $200M into the 2026 midterms
CNBC, citing Public Citizen, reports crypto-linked groups have spent more than $200 million on the 2026 midterms, the largest cohort of single-issue super PACs in US history. The Fairshake network backs 33 House candidates, 19 Republicans and 14 Democrats, with at least $1 million behind each of six incumbents. Crypto super PACs spent about $250 million in 2024.
a16z podcast: Eve says AI lets plaintiff lawyers handle 500 clients each
On the a16z podcast, Eve CEO Jay Madheswaran told general partner Alex Rampell that Eve automates plaintiff law from the first client call through settlement. According to the episode description, some firms using it have raised the number of clients each attorney manages from roughly 100 to 500. He argues the gain comes from cases that were previously too small to take, not only from lower costs.
Dazzle, Duckbill and Wajo still route AI agent tasks to humans
CNBC reports that personal-agent startups including Marissa Mayer's Dazzle, Duckbill and Wajo still rely on human workers for phone calls and other real-world tasks when software falls short. Models from Meta and OpenAI still can't reliably act across the web, especially when a task requires a phone call.
TikTok adds a restaurant guide and tests DM payments on the way to a super app
TikTok launched Must Eat, a restaurant guide built from more than 40 million posts and comments, searchable by city, cuisine, price, rating and location. TechCrunch reports TikTok has also been spotted building a way to send money in direct messages, adding to TikTok Shop, a local map and games.
Writers are changing their style to pass AI detectors, Bloomberg reports
Bloomberg reports that as detectors such as Pangram police what counts as machine-generated text, some writers are dropping em dashes, "not X, but Y" constructions and words like "delve," or adding typos, so their work reads as human.
DeepMind and WRI release ForTy, a global 10-meter forest-type map
Google DeepMind and the World Resources Institute released ForTy, a 10-meter-resolution map of 2020 land cover that sorts areas into six classes, including primary forest, plantations and tree crops. It is meant to help track deforestation under EU rules. The Decoder notes it identifies only about half of planted forest and tree-crop areas.
SpaceX's Shotwell gives 2M+ shares to low-income kids via Trump Accounts
SpaceX president Gwynne Shotwell said she and her husband will give more than 2 million SpaceX shares to eligible children through Trump Accounts, to be distributed in December. The first gift goes to 11- to 17-year-olds in zip codes with a median family income of $57,000 or less, Business Insider reports.
Scribd notifies users of breach exposing emails and partial card data
A Hacker News user posted a security notice from Scribd Inc., parent of Scribd, Slideshare, Everand and Fable. It says an unauthorized third party may have accessed email addresses, the last four digits of credit cards, card issuer, expiry date, country and payment history. The user said the notice arrived on October 26.
Researchers match 3D-printed parts to their filament in 26 of 28 cases
Curtin University researchers in Australia developed a way to match 3D-printed PLA parts to the filament product used, correctly identifying 26 of 28 samples. Police could use it to link seized 3D-printed ghost guns to their source material.