Top Stories
Anthropic’s IPO ends its first day up 17% amid an AI selloff
Anthropic priced its offering on November 18 at $83.00 a share, the company announced in its own newsroom. We covered the plan in March: Anthropic Eyes October IPO at $60B+ Valuation. The sale covers 980,000,000 shares of Class A common stock, 956,351,354 of them sold by Anthropic and 23,648,646 by selling stockholders, who receive none of the proceeds. Trading began on the Nasdaq Global Select Market on November 19 under the symbol “ANT”, the offering is expected to close on November 20, and the underwriters were granted a 30-day option on a further 147,000,000 shares. The prospectus names Goldman Sachs and Morgan Stanley.
The opening was the high point of the day. TechCrunch reports that in the early minutes shares “were trading at $111.50, a 33% premium over the $83 IPO price”, and that the opening price “would put the company’s market cap over $2 trillion, making it more valuable than Meta, Tesla, or Walmart.” The Verge reports that the $83 price itself valued the company at around $1.5 trillion. By the close the picture had changed: CNBC reports that Anthropic “closed up 17% from its IPO price” on a day the Nasdaq Composite fell 2.25% and the S&P 500 dropped 2%. CoreWeave fell more than 6%, Alphabet and Amazon also declined, and AMD closed down more than 7%.
The money is what makes the numbers worth reading. TechCrunch reports that the offering will raise over $81 billion and that the IPO, existing fundraising and new private offerings timed to the debut will leave Anthropic with an expected $220 billion in cash reserves. One analyst told TechCrunch “Anthropic’s going to have more cash on hand than any company has ever had.” The S-1 sets out a governance structure that The Verge notes keeps control of the company in the hands of the founders rather than the shareholders, and it reported an $80.7 billion net loss in the first half of the year, mostly an accounting charge, alongside $16.3 billion in revenue. It also lists possible “catastrophic” risks including “self-preserving behaviors” like resisting attempts to shut the models down. Ahead of the opening bell, Dario Amodei said, per Reuters as reported by The Verge, “To state it unequivocally, we will pace the training and deployment of AI systems, taking as long as is necessary to have confidence that they are safe and will remain under human control,” and “We will not bend to either business nor political pressure.”
The founders cashed out nearly $2 billion, and the pledge is not binding
TechCrunch reports that each of Anthropic’s seven founders sold roughly 3.4 million shares and received about $280 million before underwriting fees and taxes, nearly $2 billion collectively. Anthropic raised $81.3 billion overall and received $104.8 billion for its own coffers, potentially rising to $117 billion if bankers take up their full allotment. It also sold $26 billion of special non-voting Class N stock, with Nvidia and Google putting in $10 billion each and AMD, Broadcom, Micron, Microsoft and TSMC taking the rest between them. Google now owns 13.4% and Amazon 14.6%, a combined 28%.
The part that outlives the news cycle is the commitment attached to it. The founders pledged to dedicate 80% of their personal Anthropic equity to philanthropy, but TechCrunch reports that the filing describes those pledges as “voluntary, revocable expressions of each co-founder’s philanthropic intent” that “do not constitute legally binding obligations.” Separately, Anthropic has committed, subject to stockholder approval, 5% of its shares to a new Public Benefit Fund, with an extra 2% added at the board’s discretion for every 50% increase in the company’s value above a set baseline. Compute commitments now dwarf the balance sheet: Anthropic expects to need $70.7 billion for compute and infrastructure over the next 12 months and $451.6 billion through the end of 2030.
Meta’s Muse mixed conversations and surfaced one user’s bank balances
TechCrunch reports that Meta’s AI agent Muse is mixing up conversations and could be exposing private data, based on a Reuters investigation into internal Meta posts. Employees saw Muse conflate chats belonging to different accounts, surfacing requests about childcare, career advice, product orders, financial balances and a Signal reminder. One employee said Muse surfaced a reminder to respond to a message from her sister on Signal, without her ever granting access to other apps.
The figures in the report are specific. The balances that showed up belonged to a Capital One account holding $2,722.58 and a Charles Schwab account holding $5,899.57, and the information appeared in a conversation that was not the employee’s own. An affected employee called the incident “an extreme privacy violation.” TechCrunch reports that Muse has more than 10 million downloads. Meta’s response is that this is a testing artifact: a spokesperson said “Employee testing is core to the product development process and the entire point is to get feedback, so we can implement safety and privacy protections and improve products,” and that employees use different Muse versions and test features unavailable in the public build. That is the open question, since the distinction between an internal build and the one available publicly is exactly what the reported conversations do not settle.
Dutch watchdog says do not buy AI smart glasses
The Autoriteit Persoonsgegevens, the Dutch data protection authority, has advised the public not to buy smart glasses, the BBC reports, writing that the devices “can get things dangerously wrong.” In tests of six glasses on the Dutch market, one pair gave dating advice to an adult who indicated a romantic interest in a minor, and during a simulated emergency in which CPR advice was requested, one pair claimed several times that an ambulance was on the way when emergency services had not been contacted. The glasses also made errors reading instructions inside medicine packages and interpreting severe weather warnings.
The privacy findings matter just as much. Lights that warn a device is recording could be covered with tape and the glasses kept working, despite some manufacturers claiming they would stop. On some of the tested devices a nearby stranger could connect without permission, view recordings or operate the camera remotely, and data recorded was often exported outside the EU. The authority stopped short of calling for the products to be banned from sale and did not name what it tested, with prices ranging from under 100 euro to over 500 euro. Meta, TCL and HTC were approached for comment. The AP said the investigation was not primarily about data protection but about algorithm reliability, though it recognised the glasses can be valuable for people who need help understanding their surroundings.
The Senate passes BAAJA 62-33, easing permitting and rewriting transmission rules
CNBC reports that the Senate passed the Bipartisan American Affordability and Jobs Act of 2026 by 62 votes to 33. The bill eases environmental permitting laws, overhauls electricity transmission regulations, and includes measures to prevent cancellation of already-permitted wind and solar projects. It would also require hyperscalers to fund certain grid upgrades. Opponents say it weakens environmental protections and hands a benefit to data centers and fossil-fuel companies. The bill still has to clear the House before it reaches President Trump’s desk, and the Senate vote already split Democrats, with more of them voting against the permitting measure than voting for it.
Sen. Martin Heinrich of New Mexico argued for speed, saying “Working people can’t afford to wait another few years for us to get this done,” while Sen. Ron Wyden of Oregon put the objection in physical terms, saying “I’m not going to sit around and let these data centers scar our treasures,” about an amendment that would have blocked data centers from being built on public lands and did not pass. Sen. Ruben Gallego of Arizona argued the other way, saying “It’s going to be legislation that’s really going to make the biggest difference for the next 100 years, especially bringing down carbon emissions once this president is gone,” with the carbon argument aimed at what follows the current administration.
The AI angle is explicit in CNBC’s reporting, which says AI and the broader tech industry stand to benefit despite the provision forcing hyperscalers to pay for grid upgrades, because power-hungry data centers demand more electricity. Rhodium Group estimates the bill would allow an additional 17 to 249 gigawatts of clean energy to be built by 2040, and Grid Strategies founder Rob Gramlich said “A lot of Democrats are involved and supportive in this deal because expanding grid capacity would integrate a lot of clean energy,” with Republicans less interested in that benefit but more focused on economic growth and serving new large loads.
A study finds data centers land more often in majority-Black neighborhoods
Researchers at George Washington University and the Center for Security, Innovation and New Technology at American University examined 81,000 census tracts and location data for 5,291 data centers, 404 Media reports. The study, “Redlining in the Age of AI: The Relationship between Data Center Locations and Race in the US”, finds that a majority-Black neighborhood has 2.2 times the odds of hosting a data center than a white neighborhood in the same state, a majority-Hispanic neighborhood 2.6 times the odds, and a majority-Asian neighborhood 4.2 times the odds. The researchers write that “As the share of Black, Hispanic, and Asian residents in a neighborhood increases, so do the odds that the neighborhood hosts a data center,” and that “We found that these racial disparities persist even after accounting for factors such as population, household income, education, land costs, proximity to major transmission lines, and state fixed effects,” a result that holds even though data centers are concentrated in and around metropolitan areas, which are often more racially diverse.
The mechanism they point to is power. Data centers are often built near transmission lines because they provide ease of access to power infrastructure, and those transmission lines were often built in majority Black neighborhoods, so the researchers say the historical legacy of redlining is to blame for some of it. Bradshaw and co-author Thakur said they started from reporting on data centers in Memphis, Tennessee, and on a site in Taylor, Texas, where a white family donated a tract of land to the city to become a park, the park was never built, and the land is now slated to become a data center next to a Black neighborhood. CSINT executive director Samantha Bradshaw said “We really wanted to look at this on a nationwide scale to see: is this something that fits into this broader structural racism story?” and “What we did find was yes […] neighborhoods that have data centers tend to have a higher ratio of people of color.” Co-author Dhanaraj Thakur doubted the intent: developers were not, he said, trying to put data centers in communities of color. “It’s just how embedded race is with all these other factors,” he said. The study’s recommendations include data center-specific zoning rules and racial impact statements before sites are built, and it lands in the same week as the Senate permitting vote that decides where the next cohort gets sited.
ICE is paying AT&T analysts to sift raw phone records
404 Media reports that ICE is paying AT&T millions of dollars for access to raw phone data and for AT&T’s own analysts to examine and interpret it for Homeland Security Investigations, ICE’s criminal investigative arm. The restricted procurement document says AT&T analysts will convert the data into a usable format, and the reported capabilities include linking devices or phone numbers to a target, identifying devices whose SIM cards were replaced, determining whether a target’s phone entered the United States and deconflicting international phone numbers. The program covers more than one trillion phone records and may include location data, and because AT&T acts as a wholesale carrier, records belonging to customers of other carriers may also be included. Public records show ICE paid AT&T more than $2.1 million in 2025 and more than $2.2 million in August 2026, with potential payments exceeding $10 million over coming years.
An administrative subpoena can be issued by an agency without a judge’s approval, which is where the oversight question sits. The earlier Hemisphere program was shut down in September 2024; documents obtained by Sen. Ron Wyden’s office describe Data Analytical Services, also called Data Analytics and Support Services, as the continuation or former name of Hemisphere. Wyden said “The Biden administration shut down Hemisphere after I and Rep. Biggs repeatedly questioned this warrantless mass phone surveillance program.” Surveillance researcher Christopher Soghoian said “Hemisphere is a surveillance program that refuses to die and keeps coming back with different names and different forms,” and “It’s bad enough that AT&T is mining its customers’ private phone records as a paid service for ICE, but because AT&T is also mining the records of other wholesale telecom companies, consumers can’t vote with their wallets and switch carriers to protect their privacy.” AT&T said “We are required by law to provide information to government and law enforcement agencies when it is demanded under lawful order” and “This is not a contract supporting ICE activities, nor related to ‘surveillance’ as claimed in this inaccurate report.”
Gemini can now drive your Mac, and Apple is preparing new access rules
Google’s Gemini app for macOS now has computer use, The Verge reports, with Google saying the model can help “get multi-step tasks done across your local files, desktop apps, and the web,” and that users can decide which apps Gemini may access and follow along as it works, since it has its own background window showing a live preview in the chat. Google suggests the feature for filling in forms, making presentations and listing items for sale, and it is rolling out now in English for users in the US.
The timing is the story. The Verge notes that Google is introducing the feature ahead of Apple launching new restrictions for apps wanting “full disk access” on Mac, citing the risks from AI agents with computer use, with a preview of the settings rolling out with that week’s macOS 27.3 beta. An agent with a live view of your screen and a grant to read your files is a much bigger surface than a chatbot window, and the permission screen is the only thing standing between the two.
Apple is hiring for its own AI agent
The Verge reports that Apple is working on an AI agent that can browse the web, work across apps, write code and call services, based on a job listing spotted by Bloomberg. The role leads the creation of an agentic AI assistant and would work with the Siri, UI Frameworks and Apple Foundation Models divisions. The listing describes the target plainly: “People already do their most valuable work by moving fluidly between tools: a browser, a document, a terminal, an editor, a service,” and “We’re building AI agents that can do the same…” Those agents, the listing says, plan and carry out multi-step tasks across apps, browsers, command lines, code, and APIs, and can coordinate specialized sub-agents along the way. The Verge notes that an agent from Apple would put it in the running against Meta’s Muse, OpenAI’s Dots, Anthropic’s Duet and Google’s Tasks, and that iOS 27.3 gives developers the ability to integrate Siri AI into “all apps” as part of a Developer Preview.
Google ships Tool Retrieval so large tool libraries stop eating the prompt
Google has put Tool Retrieval into preview in the Gemini Interactions API, the company’s developer blog explains. The idea is to keep a full tool library available while loading parameter schemas only when a tool is actually needed. Developers add a tool_search tool and set defer_loading: true on the tools whose schemas should be held back, and the model keeps each deferred tool’s name and description in context while the parameter schema stays out until it requests that tool by name. The search runs server side: “There’s no index to build or host.” The feature works with Gemini 3.6 and newer, including all Gemini 4 models, and “Tool Retrieval has no separate charge.”
It also applies to remote MCP servers: setting defer_loading: true on an mcp_server tool makes every tool that server exposes load on demand. Client-side retrieval is supported as well, by setting "execution": "client" on tool_search and having the application do the registry search, which costs one additional round trip per search. Loaded schemas are billed as input for the rest of the conversation. It is the same pattern a self-hosted runner can copy for MCP servers, and what it moves is the input tokens a long tool list burns before a single tool is called.
The European Commission confirms a cyberattack carried out with AI
The European Commission has confirmed an attack on its own systems that used AI tools, The Independent reports, making it the first time the 27-member bloc has disclosed a cyberattack carried out with AI. Spokesperson Thomas Regnier said “AI-assisted tools can be used and especially misused to detect vulnerabilities,” and “Indeed we can confirm that this was done with AI tools,” adding that “We have seen many global cyber attacks recently, and of course no one is immune to this.” The EU detected the incident on 22 October, and officials immediately opened an investigation while implementing preventative measures. Regnier said “We temporarily took down several applications, and most of them have since then been restored,” and added that “we have not detected any evidence of a data breach affecting commission staff as a whole.” No related incidents were reported by the European Council or the European Parliament.
The tooling behind it has a public name. Digital security for the bloc’s institutions, which together employ around 40,000 staff in Brussels, is overseen by CERT-EU, and that agency warned on 4 November that “We assess that ARTEX’s public availability will likely accelerate adoption by threat actors, resulting in more frequent, faster, and more scalable automated attacks with a higher success rate,” a warning about a tool built to use AI to find vulnerabilities in digital networks that has already been linked to attacks elsewhere.
Quick Hits
- Canva added a switch that turns off its generative AI features, disabling Magic Expand, image generation in the Canva chatbot, Magic Write, Magic Design, Magic Layers, Magic Edit, Dream Lab, Canva Code and AI Shapes. The background remover and Magic Grab stay on because Canva says they “make everyday tasks faster without generating anything new.” A second toggle shows human-created library content first. The Verge
- ChatGPT gained Live Dictation, transcribing audio prompts as you speak instead of after you stop, with the 10-minute recording limit removed on all plans across web, desktop, iOS and Android. Tapping the waveform shows the live transcript so it can be edited in the prompt box. The Verge
- Google’s Search homepage is being rebuilt around Gemini, with a sidebar carrying chat history and tasks plus shortcuts for image creation, file analysis and brainstorming, replacing the existing Search button. Google’s Robby Stein says Gemini gets near-top billing on the page, and the redesign is rolling out globally on desktop. 9to5Google
- Superhuman’s Go assistant is going to individual users, handling personal email and calendars on top of enterprise search and meeting context, with CEO Shishir Mehrotra saying it is past 1 million active users. TechCrunch
- Google built a controlled replica of the internet to train Chrome’s browsing agents, forging web certificates so agents could pass real Chrome validity checks and facing them with hidden prompt-injection attacks. The team says none of those attacks would work against regular users on the open internet. WIRED
- AMD closed down more than 7 percent after a short-seller report alleged its firmware turned up in US-blacklisted companies, and House Foreign Affairs Committee documents show a Commerce Department subagency approved sales of more than 879,000 AI chips to Chinese companies. AMD said it complies with US law. CNBC
- An AI program called Caucus Pulse was built and run by a staffer in the House Democratic Leader’s office, driven by Claude and Claude Code, with its GitHub repository built using Claude plus ChatGPT, and told it was “a research analyst in the House Democratic Leader’s research and writing shop, writing for staff who have thirty seconds.” It read 114,647 posts and viewed 69,285 accounts while tracking caucus messaging. The office said the leader was not aware of it and that no official resources were used. The Washington Sun
- Free-trial abuse has become a business model for AI agents, with groups looping trials and promo codes to resell tokens or run commercial work on someone else’s free tier. ChatPRD founder Claire Vo said “I am funding a international Roblox and Fortnite map studio” and is winding down free trials. Simon Willison
- Rumble-2-Flash, an open 0.5B speech recognition model for 40+ languages, was released under Apache 2.0 on November 18. Its authors say it competes with models 3 to 60 times its size, transcribes an hour of audio in about half a second on full batches, and matches NVIDIA’s 30B Nemotron 3.5 Transcribe on a 302-hour call-centre benchmark. Hugging Face
- Streaming MoE experts straight from SSD now lands within 6 percent of Strata on token generation, a November 18 Hugging Face post reports, with one example running Qwen3.8-Flash-Next on a 32GB Mac at 13.5 tokens per second. Hugging Face
Worth Watching
- Whether the probable-cause ruling survives. In Osorio v. Kilmer, a federal judge in Kentucky held that the Fourth Amendment does not let a detective delegate the probable cause determination to AI, and declined to give qualified immunity to the detective at the pleading stage, so an arrest resting only on an uncorroborated AI face match can proceed to discovery. ID Tech reports the case is Osorio v. Kilmer, No. 2:26-cv-00004, before Judge S. Chad Meredith in the Eastern District of Kentucky. Biometric Today reports the order was signed on November 9, 2026 and that a status conference was set for November 19, 2026. Watch whether other courts follow the reasoning on warrant affidavits.
- How Apple’s new macOS access rules land. A preview of the settings for apps requesting “full disk access” shipped with the macOS 27.3 beta, in the same window that gave Gemini live control of the desktop. If the restrictions hold, they set the template for how much of a machine a desktop agent gets by default. The Verge
- Whether BAAJA survives the House, and what it does to grid buildout. The permitting and transmission rewrite now has to clear a second chamber, and the hyperscaler grid-upgrade funding requirement is the provision most likely to decide how fast new AI capacity can be connected. CNBC
- Who actually holds Anthropic. Nvidia and Google each took $10 billion of non-voting Class N stock, and with Google at 13.4% and Amazon at 14.6%, a company that says it keeps control with its founders is now partly owned by the largest AI infrastructure vendors and its main rival. TechCrunch
- Whether the AI trade reset sticks. Anthropic closed up 17% while the Nasdaq Composite fell 2.25%, CoreWeave fell more than 6% and AMD more than 7%. Options trading on Monday is the next test of whether this is a one-day flush or a repricing. CNBC
Related on Intelligibberish
- AI Privacy Audit: What Eight Popular Assistants Collect - what apps gather, and how the Muse story fits the wider pattern
- Silicon Valley’s Shadow Power Grid: AI Builds Its Own - the grid buildout behind today’s permitting vote and siting study
- News - the daily news hub for the rest of the November 2026 roundups