Binance has confirmed to TechCrunch that it is dealing with a security incident as users reported that funds were being drained from the crypto exchange’s wallets.

The world’s largest cryptocurrency exchange said in an emailed statement to TechCrunch that it is “aware of reports of a potential security issue and are actively investigating.”

“Rest assured that user funds are completely safe,” Binance said in a statement.

“Our security teams are reviewing all relevant information with urgency, and we will provide updates through official Binance channels as soon as we have verified information,” the statement added. “Users should remain cautious of scams and should not trust unofficial links, messages, or requests for account information.”

Binance spokesperson Ross O’Leary said that Binance has “temporarily suspended withdrawals on multiple networks following a security incident as a precautionary measure.”

News of the breach comes hours after Binance founder Changpeng Zhao said in a post on X that the company had been likely hacked. Zhao, who no longer runs the company but is its largest shareholder, said that he expects users’ funds to be “fully covered,” but did not provide an explanation for the claim.

Financial tech giant Circle froze about $800 million in Tether as it passed through its networks, while crypto giant Tether itself froze about $1.56 billion related to Binance’s hack.

Crypto watchers say at least $2 billion has been drained from Binance’s hot wallets, which are internet-connected cryptocurrency wallets that store crypto that are connected to the internet. By some estimates, Binance may have lost upwards of $5.7 billion in crypto, though the numbers have not been confirmed and it’s possible that the theft is ongoing.

It’s unclear who is behind the hack, and Binance did not say if it had received any correspondence from the hackers.

If confirmed into the billions, this would likely rank as one of the largest crypto breaches in history, according to leaderboards tracking crypto thefts.

In 2025, hackers stole a record $1.4 billion in crypto in a single breach from exchange Bybit, blamed on North Korean hackers.

In what appears to be a separate incident earlier on Thursday, prediction market giant Polymarket briefly halted trading after it confirmed that a third-party company was compromised, which led to the company pausing withdrawals. According to security researchers, hackers stole around $2 million in user funds after compromising a piece of software developed by a company called DebugBear.

In Brief

Russian president Vladimir Putin said that he and SpaceX CEO Elon Musk are in “close contact” about Rassvet, Russia’s answer to Starlink, Musk’s internet satellite service.

“Elon Musk is a genius, truly a genius,” Putin said at the Russia-Africa Summit on Wednesday. “He has already had more than 5,000 such satellites in orbit, while we are only taking the first steps.”

A potential relationship between Putin and Musk could have serious implications. Musk has recently voiced frustration with Ukrainian president Volodymyr Zelenskyy, whose country has been fighting a Russian invasion since 2022. After the U.S. eased sanctions on Russia earlier this month, Ukrainian officials became concerned that Ukraine could lose access to Starshield, the military version of Musk’s Starlink network, which Ukrainian forces rely on.

There was some merit to these concerns. For about 10 days earlier this month, Ukrainian drones experienced frequent losses of connectivity. According to the Wall Street Journal, Musk ordered SpaceX to limit Ukraine’s Starlink access, though other reports say the order came from the White House. Musk claimed in a post on X that Ukraine violated Starlink terms of service by using the technology for “offensive drone strikes to kill people.”

Starlink service in Ukraine appears to have since resumed normal operations.

“Were Starlink to be turned off, the Ukrainian front line would collapse immediately,” Musk also wrote on X. “This is NOT a threat, it is simply a statement of fact.”

Amazon reported its third-quarter earnings on Thursday and missed analysts’ expectations for both its total revenue and earnings per share. There were other murky numbers in the report that scared investors, too, including how much cash the company is burning through.

But the report did have a big, bright spot: Its cloud unit, Amazon Web Services, beat revenue expectations and grew a whopping 39% compared to the year-ago quarter.

Here are the numbers:

Net sales increased 12% to $201.9 billion. But analysts were expecting $202.2 billion, according to FactSet.

Fully diluted earnings per share were $1.89. Analysts were expecting $1.97.

AWS sales reached $45.98 billion. Analysts were expecting $45.29 billion. In addition, Amazon said that AWS operating income grew to $17.8 billion, compared with $11.4 billion in the third quarter of 2025.

But here’s the biggest ouch of the report: Free cash flow fell to a negative of $42.4 billion for the trailing 12 months — meaning the company was spending far more cash than it generated. This compares to positive free cash flow of $14.8 billion for the trailing 12 months ended September 30, 2025.

The big spending was driven largely by Amazon’s investments in AI, the company said. With one quarter left in its fiscal year, Amazon reiterated that it still expects to end its 2026 with cash capital expenditures of about $220 billion.

To put that in perspective, the company forecast fourth-quarter net sales of between $240 billion and $247 billion. The fourth quarter is typically a retailer’s biggest, thanks to the holiday shopping season. That means Amazon expects to spend almost as much on capital expenditures, largely for AI, over the full year as it expects to bring in from sales in the fourth quarter.

General Motors will pay $220 million to 36 states to settle allegations that it illegally sold private driving behavior and location information to data brokers.

As part of the settlement, GM must get drivers’ express consent before it collects or shares their driving and location data — terms that are in line with separate agreements it previously reached with California and the Federal Trade Commission (FTC). This settlement does contain a new requirement that GM must have a third party assess the automaker’s compliance with the privacy program it has put in place.

The agreement announced Thursday settles a lawsuit brought by a coalition of attorneys general from 36 states that alleged GM had sold the information to data brokers, which in turn then sold it to insurance providers.

“This agreement addresses Smart Driver, a product we ended in 2024, and reinforces steps we’ve taken to strengthen our privacy practices,” GM said in an emailed statement. “Safer, more helpful and personalized experiences are enabled by connected vehicle data, and GM is committed to being clear about what information is used, how it’s protected, and the choices customers have about what they share.”

The settlement comes more than two years after a New York Times report showed how GM and OnStar collected, used, and sold drivers’ precise geolocation data and driving behavior to third parties, including data brokers LexisNexis and Verisk. The data was collected through GM’s Smart Driver program, a free feature contained within its connected car apps that tracked and rated driving behaviors and seat belt use. Data brokers sold this information to insurance providers, which could affect customers’ rates, the NYT reported at the time.

GM discontinued Smart Driver in 2024, but that didn’t stop legal action. In January, the FTC finalized an order that bans GM and its OnStar telematics service from sharing certain consumer data. The FTC settlement, as well as the multistate one announced Thursday, also requires GM to be more transparent with its consumers and receive their explicit consent for certain collection.

The company now says it is focused on protecting the privacy of its customers and will no longer sell vehicle information that can be linked to an individual, their vehicle location, or how they drive. However, GM may still disclose anonymized data not tied to an individual with certain partners. GM has also updated the language in its privacy policy and created a consumer-facing Privacy Center.

The settlement and mandated changes to how GM handles customer driving data come as more sensor-laden vehicles equipped with computing power enter the market. Automakers have made large bets on software-centric vehicles that drivers and passengers can interact with and connect with other digital apps such as their calendar, messages, and email.

For instance, GM is adding a personal AI assistant to OnStar, its 30-year-old safety and roadside assistance service. Cameras on the vehicle will be connected to the software to offer live views as well as crash and safety recordings. Users will also be able to use voice to control certain functions on the vehicle, including climate, and ask questions about any specific vehicle feature.

The automaker emphasized its privacy protections when it announced the new assistant during its GM Forward event in October. The company said customers will have the ability to decide what data is shared and that the recordings have time limits and are encrypted. GM said it and OnStar cannot access Connected Camera recordings. Customers will also be able to manage permissions. However, customers who opt not to share their data will likely have limited access to certain features, according to GM.

In Brief

Archive of our Own (AO3), the popular site where fans publish and read stories about their favorite characters and worlds, has bad news for its community. After going offline for what was supposed to be routine overnight maintenance, the site says it could stay online for almost a whole month.

“We estimate that service restoration for AO3 will take around 2 weeks but it could be up to 4 weeks,” the AO3 status page says. “We are hopeful that we will be able to bring the site back up in a degraded state sooner.”

AO3 is a project of the nonprofit Organization for Transformative Works (OTW), which relies on volunteers and donations to keep the site afloat. The site has such a passionate user base that, within two hours of the announcement, its status page’s posts had garnered millions of views (and much panicked commentary) across social media.

We estimate that service restoration for #AO3 will take around 2 weeks but it could be up to 4 weeks. We are hopeful that we will be able to bring the site back up in a degraded state sooner. Refer to our status page for updates: https://t.co/YAyyoTDOpE https://t.co/PUWaHF9M8L — AO3 Status (@AO3_Status) October 29, 2026

Some fans may suspect this issue is related to last week’s “PHP Armageddon,” when tools from OpenAI and Google Wiz uncovered two critical flaws in PHP, the programming language that runs the behind-the-scenes software of about 70% of the internet. But AO3 is not among that majority. Instead, the site is built on a different framework (a toolkit for running websites), Ruby on Rails, so the outage is unlikely to be related.

OTW and AO3 did not immediately reply to requests for comment. We’ll update this article if we learn more.

Reddit has long attracted millions of users to its site through Google, with many even appending the word “Reddit” to their search queries for better answers. But as search traffic dries up amid the shift to AI-powered search, Reddit is rethinking its relationship with the open internet. In its third-quarter earnings call on Thursday, the company announced it’s transitioning to a primarily “logged-in experience.”

The move is designed to protect its data from AI scraping, allowing the company to instead forge lucrative licensing deals with model providers. But just as important, such a shift could help Reddit to increase its revenue even as it faces daily active user declines in its largest market, the U.S.

Overall, Reddit’s global daily active users were flat quarter-over-quarter in Q3, reaching 130.3 million, while U.S. daily active users dropped from 53.2 million in Q2 to 50.7 million in Q3 and were down 2% year-over-year.

Image Credits:Reddit

The company is not alone in seeing such declines; the shift to an AI-centered web has impacted most websites and publishers by now, leading to fewer click-throughs, as AI summaries provide answers that don’t require website visits and Google’s traditional “ten blue links” have reduced relevance. (The change is now even impacting Google’s bottom line, its earnings revealed.)

But at Reddit’s scale, the numbers associated with such a decline are enormous. The company said search traffic referrals dropped by more than 1 million referrals on average per week in October. It warned that if the trend continues, daily active users could decline to a low-double-digit percentage by next quarter. (To play that out: a hypothetical 10% to 12% decline could bring Reddit to roughly 115 million to 117 million daily actives, down from 130.3 million in Q3. That’s a huge drop!)

At the same time, Reddit has gotten significantly better at monetizing its audience even as its growth weakens. Revenue increased 53% year-over-year to $895 million, ahead of Wall Street’s expectations of about $875 million. Net income rose 74% to $283 million. Advertising revenue grew 54% to $847 million while nonadvertising revenue, which includes its AI data-licensing business, grew 35% year-over-year to $48 million.

And despite a 2% year-over-year decline in U.S. daily users, Reddit’s average revenue per daily active user in the U.S. rose 53% to $13.82. Globally, average revenue per daily active user increased 36% year-over-year to $6.87.

Shifting to a logged-in experience could boost those numbers even higher because logged-in users are typically worth more to advertisers.

Unlike the logged-out users who often find their way to the site via search, logged-in users provide more signals about their interests and behaviors, which helps to better target and personalize ads.

Those same signals would also help Reddit provide a more personalized experience to its users, which includes personalized feeds, recommendations of new communities to subscribe to, and pushes to vote, comment, participate, and more. (This also includes pushing recommendations to its new algorithmic “For You” feed, announced yesterday, which will combine posts from communities users belong to alongside other recommendations based on their interests.)

Image Credits:Reddit

Still, a change of this magnitude could be the canary in the coal mine for the open web.

“I don’t know about you, but as someone whose life has been defined by the open internet, it’s a big change for me,” wrote Reddit co-founder and CEO Steve Huffman in a post published on Reddit’s site. “But I’ve come to the conclusion that this is the right decision for our platform and the right decision for our business,” he said.

The company notes that the shift to the logged-in experience will not be an overnight change; it will first run tests in select markets outside the U.S. in the fourth quarter. It doesn’t yet have a timeline for when the change will be rolled out to all its users worldwide. It also acknowledged that access to a logged-out front page is important.

Plus, Reddit pointed out that being logged in doesn’t mean having to give up anonymity, nor will it remove its option to browse anonymously through its “incognito mode” experience, which is being rebuilt.

“There are still plenty of questions that are unanswered, particularly around where the boundaries should be between logged-in and logged-out experiences,” Hoffman wrote. “This is a big evolution for Reddit, and not one we’re taking lightly. But I believe it’s the right next step for the platform and for our future.”

The $5.8 billion joint venture between Rivian and Volkswagen Group is expanding to include AI assistants.

Rivian is now licensing elements of its AI voice assistant technology to Volkswagen Group as part of the broader joint venture, according to the company’s third-quarter investor presentation released Thursday.

Rivian didn’t disclose the licensing fees, payment terms, or which Volkswagen vehicles will use the technology. However, the AI assistant, which recently rolled out to Rivian vehicles, could make its Volkswagen debut in the automaker’s low-cost EV called the ID EVERY1. The small four-door hatchback is slated to be the first VW vehicle to use Rivian’s software and electrical architecture under the joint venture established in 2024.

The joint venture between Rivian and VW Group has already expanded once before, at least financially. The companies initially announced a joint venture valued up to $5 billion, but within months VW upped its commitment to invest up to $5.8 billion in Rivian by 2027. The initial tie-up, formalized through an independent company that kicked off in November 2024, focused on bringing Rivian’s electrical architecture and software to Volkswagen’s next generation of vehicles.

Rivian described the expansion to include elements of its AI assistant technology as “a proof point for the extensibility of the platform we’ve built in-house.”

Rivian spent two years developing its AI assistant before showing it off last year. That development work continued longer than expected, with the assistant finally rolling out in May 2026 to its flagship R1S SUV, R1T pickup truck, and in October to its new midsize R2 SUV.

Rivian has big aspirations for its AI assistant, which allows drivers and passengers to interact with the vehicle and control certain functions. Rivian founder and CEO RJ Scaringe told TechCrunch onstage at Disrupt 2026 that the company has created its own in-house agentic framework that understands what action a person wants to take when they talk to the vehicle, and routes it to the right part of the car’s software. That capability could eventually allow Rivian vehicles to run AI agents created by other companies, including Anthropic.

The in-house AI assistant program also fits into Rivian’s push to become more vertically integrated — a strategy the company emphasized in its third-quarter earnings report.

“Our relentless focus on vertical integration remains a core competitive differentiator, allowing Rivian to seamlessly unify our hardware and software architecture to deliver significant value to our customers,” the company noted in its investor presentation.

That strategy led to Rivian overhauling its flagship R1T truck and R1S SUV, changing everything from the battery pack and suspension system to the electrical architecture, sensor stack, and software user interface. Its new electric architecture system and compute platform slashed the amount of wiring used and reduced the total number of electronic control units, or computers, used to control the vehicle from 17 different ECUs to seven. That foundation, Rivian argued, would allow it to build vehicles that can handle the computing demands required for advanced software and automated driving.

That technology caught the attention (and money) of VW, which has struggled to offer customers a modern software-centric vehicle found in competitors like Tesla.

Rivian made that vertical-integration case repeatedly in its investor presentation, noting that because it controls the entire technology stack, including compute, sensors, displays, lighting, and sound, “it can roll out rapid, vehicle-wide OTA updates that go well beyond the basics.”

This article was updated to provide the two separate rollout dates for the AI assistant.

Rivian’s new R2 SUV was its best-selling vehicle in the third quarter, the company said Thursday, marking an auspicious start for the all-important EV.

While Rivian didn’t say exactly how many R2s it sold over the three-month period, the company is confident enough in the pace of sales that it raised its delivery guidance for Wall Street. Rivian had previously expected to deliver between 65,000 and 70,000 EVs this year. On Thursday, it bumped up that floor to 67,000.

It’s the second time Rivian has raised its delivery guidance this year, and the company is now on track to break a three-year stalemate during which it struggled to ship more than 50,000 vehicles annually. The company launched the R2 just in June, but it has made a major bet on scaling production of the lower-cost SUV at a historically quick rate.

The early success of the R2 is also a sign that Rivian is finding new customers despite last year’s loss of the $7,500 federal EV tax credit. The death of that tax credit, along with changes made by the Trump administration to U.S. regulations that supported the growth of EVs, has caused many major automakers to shift their attention back to internal combustion and hybrid vehicles — leaving room in the market for pure-EV companies like Tesla and Rivian.

“In a U.S. automotive marketplace starved for high-quality EV choice, R2 is an attractively priced option for everyday adventures that is resonating with a broad set of consumers and we expect will be a game changer for Rivian’s long-term growth and profitability,” the company said in a letter to shareholders released Thursday.

Rivian revealed earlier this month that it built 19,751 vehicles in the third quarter and delivered 19,248. That was roughly 45% more than it delivered in the same period last year and a big jump from the 12,613 vehicles built and 12,194 delivered in the second quarter of 2026. The increase was driven in large part by the company launching a second production shift at its factory in Normal, Illinois.

Rivian said Thursday that the R2 was responsible for “much of the gain quarter-over-quarter,” but didn’t go into specifics. The company generated $2.2 billion in revenue in the third quarter and suffered a net loss of $913 million — down from $1.2 billion in the same period last year.

Rivian launched the R2 at a starting price of just under $60,000. A more affordable “premium” trim is supposed to be available later this year that starts at $53,990. Even cheaper versions are slated for 2027, which will be available for as low as $44,990.

In Brief

Thrive Holdings has acquired physical security company ADT for $10 per share or about $7.6 billion in equity, the companies announced on Thursday. Beyond equity, the total enterprise value of the deal is $15.3 billion, ADT said.

Thrive Holdings is the PE spinout from Josh Kushner’s powerhouse VC firm Thrive Capital. It also counts OpenAI as its partner, as well as an investor. The firm’s mission is to hold other companies long-term, revitalizing them with AI. Currently, Thrive Holdings holds over 75 companies, it says.

As a partner, OpenAI supplies AI models and tech as well as people, otherwise known as forward-deployed engineers, to Thrive’s companies.

Neither ADT nor Thrive announced what new AI-powered security products they will be releasing, but they did tease that the plan was to invest and release new products. “ADT will form the foundation of Thrive Holdings’ smart home security business,” the press release said.

AI-powered physical security is having a moment right now. Apple just entered the field with a new line of smart home and home security products in collaboration with LG and Schneider Electric. Amazon updated its smart home line, including new AI-powered home cameras. And who knows what the future might bring? Some startups are offering AI dog robots as roving security guards.

Airbnb’s nonprofit held its inaugural gala on Wednesday in Los Angeles, and the internet had one reaction — the what gala?

realizing i’m the first person in my bloodline to read the words ‘airbnb gala red carpet’ one after another



they’re just making up events for anything now https://t.co/wwRDmA7Wfo — bubs!! (@etoilepattinson) October 29, 2026

In case you missed it, back in 2020, Airbnb launched a nonprofit that provides free and temporary housing to families impacted by crisis or disaster. Its gala makes sense, actually, in that significant causes often throw splashy events to please big donors.

In a press release shared with TechCrunch this week, Airbnb noted that the gala marked the launch of a new initiative for the nonprofit, in which Airbnb will start donating a portion of its commercial revenue from each stay booked starting from the day of the gala, October 28. The cut amounts to 10 cents, to be exact, but Airbnb did make $12.2 billion in revenue last year.

This initiative will help generate $500 million that will then go to the nonprofit over the next five years to help it provide free housing for 1 million nights each year, per the release. It will also let the company expand to supporting families also facing displacement due to domestic violence, veterans in need, or those traveling for medical needs.

A spokesperson for Airbnb said that the night was for its supporters and “marking the next chapter of our work.” Airbnb said that people had the option to donate, but “fundraising wasn’t the focus of the evening” and that the nonprofit “didn’t set a fundraising target.” The idea is for the Gala to happen each year, and last night was invite-only, the spokesperson said, with no tables or seats sold.

The gala was hosted by Jimmy Kimmel, with a live performance by Charli XCX, and food was curated by Chef José Andrés of World Central Kitchen. The guest list featured Serena Williams, Demi Lovato, Charlize Theron, and Cynthia Erivo. For the celebrities, it was a glamorous night in society for a helpful cause, joining the ranks of other Los Angeles events, like the Baby2Baby Gala and the Academy Museum Gala, as a place to be seen.

Still, people were literally confused as to why Airbnb was having a gala and how it attracted top-tier celebrities.

How the hell does Airbnb has a gala 😭 https://t.co/Kx7TUiyvfa — Gabriel Divina (@GabrielDivina2) October 29, 2026

If you told me literally one day ago that A-list celebrities would be attending an Airbnb Gala I would say what the FUCK are you talking about https://t.co/URgwQnSgVG — Meech (@MediumSizeMeech) October 29, 2026

the concept of an airbnb gala https://t.co/Q5LdCrGHLw pic.twitter.com/q2WegX3D5a — selina meyer’s wife (@selinameyersgf) October 29, 2026

Absolutely stunning. What the FUCK is the Airbnb Gala. https://t.co/sWexvP6fou — #1 angel (@BLAMElTONYOU) October 29, 2026

i went to the airbnb gala and everyone knew you https://t.co/DBiakwlQV4 — Daniel Spielberger (@quepaso_daniel) October 29, 2026

Again, a lovely event, and another data point showing that throwing a gala is one of the best ways to get rich people in a room to pay money and get media attention. And now a bunch of people on the internet know that Airbnb helps people in need, rather than perpetually making the news for negatively impacting the housing market.

This piece was updated.

Meta has quietly launched Style Sketch, an AI fashion app that helps users put together outfit ideas using clothes they already own. The app connects to Instagram to build a virtual closet from photos you’ve posted, or lets you add items manually by snapping a photo, recording a quick video, or uploading images from your camera roll.

Style Sketch then suggests outfit combinations built from your wardrobe matched to your style. Users can then try on the outfits on a “realistic model matched to your body shape and skin tone,” the app description reads.

The app, which was first spotted by social media consultant Matt Navarra, is available in Indonesia, Singapore, and Thailand on iOS and Android.

Meta did not respond to TechCrunch’s request for comment. It’s unknown whether Meta plans to launch Style Sketch more widely.

🚨 Meta quietly launched an AI fashion app! 👗



Style Sketch connects to Facebook & Instagram to build a wardrobe from photos you’ve already posted



• 👕 AI outfits from clothes you own

• 🪞 Virtual try-ons matched to your body

• ☀️ Weather-based looks

• ✈️ Trip packing plans pic.twitter.com/uNYX6VHRCn — Matt Navarra (@MattNavarra) October 28, 2026

Several apps already use AI to help users digitize their wardrobes and visualize new outfits, including startups like Alta. Even Google Photos offers tools that let users turn photos of their clothes into digital closets where they can create new outfit ideas.

To compete with what’s already on the market, Meta will likely rely less on the novelty of the idea and more on the app’s integration with Instagram, where users already have a visual record of their personal style.

Rather than relying on random suggestions, users can also select a specific item from their wardrobe and get fresh outfit ideas built around it, according to the app description.

The app can also help users plan outfits for trips. If they enter their destination, trip length, and planned activities, users can get a day-by-day outfit plan that takes the weather forecast and itinerary into account.

Style Sketch is one of many AI apps from Meta in recent months. Most recently, the company launched an app called Waddle World, which essentially looks like a Tamagotchi clone, but with AI. Meta has also launched Pocket, a tool for vibe-coding games, and Vibes, which lets users quickly generate AI-powered videos.

The wave of new apps is part of a broader effort at Meta to release more apps. The Wall Street Journal reported earlier this year that CEO Mark Zuckerberg told employees the company plans to release far more apps than it has historically, as AI-driven efficiencies have made it faster and easier to ship new apps.

In Brief

Apps in the consumer voice AI world seem to be converging, as Granola, known for its meeting notetaker, launched a dictation feature today. This comes months after Wispr, known for its dictation tool, launched a meeting notetaker.

Granola’s dictation currently works on Mac just like any other such tool. Users can press the “Fn” key and start speaking, and the tool cleans up and formats their speech and posts the final text. This feature works across all apps, so users can post notes, dictate prompts, and reply to emails and messages.

Granola now does voice typing. Unlimited words, for free.



Granola already writes your meeting notes. Now it can do the typing for you, in any app. pic.twitter.com/et60VtwWA0 — Granola (@meetgranola) October 29, 2026

The company said that the feature will land on iOS and Windows soon.

Granola and Wispr are valued at $1.5 billion and $2 billion, respectively. Granola is focused on meeting-related automation and notetaking. On the other hand, Wispr is trying to build its own models through a research lab, which is exploring different human-computer interfaces.

The feature mashups are not exclusive to Granola and Wispr. Other voice startups such as Fireflies, Speechify, and Monologue have converged into multifeature apps with meeting notetaking and dictation capabilities. Until earlier in the year, the play was around having a focused app that does one thing well. But now, all of these apps are thinking about how to get more out of capturing users’ speech.

French AI company Mistral has released the weights for its trillion-parameter open source AI model Mistral Large 4 (ML4), allowing businesses and governments to download and deploy the system themselves. Nicknamed “Le Chonk,” the model was first introduced earlier this month.

The release comes as competition between open-weight AI developers like Mistral and closed-model providers like OpenAI and Anthropic heats up, with the former arguing that businesses and governments shouldn’t have to rely on outside AI companies for access to advanced capabilities. That argument may hold more weight in cybersecurity, where organizations want more control over their data and systems.

Or, as Mistral’s CEO, Arthur Mensch, explained, AI should “be a tool for empowerment, not dominance.” With Chonk, Mistral doubles down on that position, arguing that it’s able to achieve “state-of-the-art performance” at a “fraction of the compute and capital committed by closed labs.”

The release of the model’s weights — the numerical training that animates the system — ensures that other companies and government agencies will now be able to deploy ML4 for their own purposes.

Mistral says that its model is comparable in terms of performance with a Chinese open-weight competitor, GLM-5.3, which was previously considered one of the best open-weight models in the world — particularly when it comes to cyber work. However, Mistral’s model uses 47 percent fewer tokens per task, making it significantly more economical than GLM-5.3.

The company also claims that the model outranks all other open models when it comes to agentic coding tasks on Artificial Analysis’ Coding Agent Index, an AI benchmarking site.

“Amid growing cyber threats, ML4 brings state-of-the-art sovereign cyber defense capabilities to enterprises and governments globally,” the company said Thursday. “Enterprise and public sector institutions can now download, deploy, and own frontier AI to detect, analyze, and respond to threats across code and IT infrastructure.”

The weights for ML4 are being made available on Hugging Face and via its API.

Pierre Stock, Vice President of Science at Mistral, told TechCrunch that the release of the model’s weights meant that organizations will have more control over how they use AI to defend themselves.

“Concretely, it means that, for instance, as a business company or as a government, you’re able to have access to frontier intelligence to defend yourselves against cyberattacks. Access that can’t be cut off by definition because the model is open-weighted,” he said.

Stock added that the model could be used to “identify, coordinate, and respond in real time to major security incidents, and that the open sourcing means that companies can deploy this model on-prem in their own ecosystem.

While advances in AI cybersecurity are helping security professionals to find bugs at a much higher rate than before, Stock continued, he believes it’s important for companies to “provide [this same] capability to everyone…as soon as possible to ramp up and to strengthen the ecosystem,” given that “some malicious actors might try to actually leverage their own technology to attack.”

Separately, Stock addressed concerns about recent incidents involving rogue AI agents, acknowledging the risks but cautioning against overreacting.

“We don’t want to fearmonger people. We think it’s a problem that is solvable. It’s a hard problem to have good security,” he said.

In Brief

A top executive of a Singapore AI cloud firm that was accused of illegally selling Nvidia chips to China could also have been a Chinese spy, anonymous sources told the New York Times.

The executive, Alice Huang (or Huang Le), worked for Megaspeed and was photographed as one of multiple people hanging out with Nvidia CEO Jensen Huang during various business dinners in the Singapore region. U.S. intelligence agencies had been surveilling Alice Huang to learn more about her ties with China for about a year, the Times reports.

The concern, sources said, was that Alice Huang had grown so close to Jensen Huang that she could have been gathering information about Nvidia’s plans and the CEO’s efforts to convince the Trump administration to allow it to ship more advanced technology to China.

But an Nvidia spokesperson denied this, telling TechCrunch that the CEO’s relationship with the Megaspeed executive was the same as with any other regional customer.

“Jensen’s interactions with Ms. Huang are not unusual or unexpected. Meetings with customers are part of a C.E.O.’s job; Jensen talks with thousands of customer executives every year,” the spokesperson said in an emailed statement. The company also said, “Our regional cloud customers receive ordinary business information — not sensitive NVIDIA trade secrets — and zero information relevant to national security.”

In 2025, the Times reported that the Commerce Department had investigated Megaspeed over concerns that it was helping Chinese companies sidestep AI chip export regulations by reselling restricted chips to them. Alice Huang was said to be a top executive for the firm at the time.

Andy Elder and Arjun Mathur were Stanford roommates who both studied computer science before spending a decade at Palantir. Elder led product design for a team of 800 people building software for U.S. defense and intelligence agencies, while Mathur served as a senior vice president in the company’s commercial practice. When the friends decided to build a startup early last year, there was no question they would do it together.

Despite having no direct background in architecture, Elder and Mathur were troubled by a growing problem: Fewer buildings are constructed in the U.S. today than a few decades ago, as development costs have surged.

“We’ve gone from dial-up internet to superintelligent AI, but our ability to build the physical world around us has actually decreased,” Mathur told TechCrunch.

Believing AI can boost productivity across the architecture, engineering, and construction (AEC) industry, they set out to build Dream, a startup developing built world models that automate building design and engineering. Dream has emerged from stealth with $17 million in seed funding led by Andreessen Horowitz, alongside other backers, including SV Angel, XYZ, and Zigg Capital. The startup aims to replace cumbersome, outdated AEC tools like Autodesk’s AutoCAD and Revit.

If you sit next to an architect or a structural engineer, you’ll notice that they are still “drawing” every piece of their designs part by part on legacy software, Mathur said. “AI really plays no role in that equation.”

While a standard LLM like ChatGPT can generate an image of a building, the result won’t be structurally sound. Dream, by contrast, is grounded in real physics and construction standards. Just as LLMs learn from text, Dream’s world model is trained on more than 100 million square feet of building data gathered from permit archives across the United States.

“Our model actually understands what’s within those walls. What materials are used, what the beams are made of, and how much weight they can support,” Elder said.

The startup calls its approach “structural cognition” because its model generates structurally precise designs that account for internal building elements like plumbing and HVAC.

Dream has already begun selling its product to AEC professionals and developers, who use the model to conceptualize building designs and estimate overall costs, according to the founders. The platform is already deployed across 10 active projects in Texas and California.

While Dream isn’t the only company attempting to update AEC software with AI — other startups working on this include Motif, Snaptrude, and Arcol — Elder and Mathur claim that they are the only ones that developed a world model that instantly designs whole buildings.

Kimberly Tan, a partner at Andreessen Horowitz, said that while the firm first invested in Dream over a year ago, when the startup model was just an idea, she has been extremely impressed with their progress. “I have not seen anybody take the technical approach,” she said. “They’re doing everything based on their own models.”

When people have career or job questions, they usually tap into their network to find experts who’ve faced a similar predicament and can offer answers. Increasingly, that job is falling to AI.

SuperMe is a network of AI profiles built from the knowledge of real-life experts. When you ask a question, the platform determines which experts are best suited to answer and creates a group chat with their AI profiles.

To scale that network on both sides, meaning more users asking questions and more experts supplying knowledge, the startup announced today that it is raising $20 million in funding led by Kirsten Green at Forerunner, with participation from Greylock Partners.

Casey Winters and Ludo Antonov founded the startup in 2024. Both have experience at large consumer companies. Winters held exec roles at Grubhub, Pinterest, and Eventbrite and has advised companies like Airbnb and Canva part-time. Antonov was head of engineering at multiple startups, including Pinterest, Lyft, and Whatnot.

Image Credits:SuperMe

The two got connected when Winters was advising Whatnot. ChatGPT was becoming popular around the same time, and they decided to build a network business using AI.

“There are these hidden networks in Silicon Valley of who the best people are for niche skills. Our idea was: Could we make that network of who the absolute best people are more transparent, and could we take all of the key insights that are locked inside their heads and make them more readily available?” Winters told TechCrunch over a call.

Users can go to the SuperMe website and type their question into the chatbot to start a group of experts. They can also return to a group they have formed or joined and ask follow-up questions.

You can ask a question like “Analyze my pitch deck. Which parts of the story are the weakest,

and which questions should I be most prepared to answer?” and get answers from AI profiles of longtime VCs like Mercedes Bent, Sarah Tavel, and Josh Elman. Users can also go to these people’s profiles and ask questions directly rather than in a group chat.

To collect knowledge from users, SuperMe scans public sources such as LinkedIn, YouTube, Substack, and blogs. It also draws on connected apps such as Google Drive, Notion, Gmail, and meeting notetaking tools like Granola. SuperMe said some of this data will include personal information and trade secrets, but it makes sure that none of it appears in answers.

Image Credits:SuperMe

The network is currently free, but the startup plans to cap the number of questions users can ask, after which they will have to cough up some money. It is also considering how to give experts incentives to add more knowledge.

SuperMe also works with MCP (Model Context Protocol, a standard that lets AI chatbots connect to outside tools and data), so users can ask questions from within chatbots such as Claude and ChatGPT.

The startup has about 13 employees. It is hiring on the tech side to improve its AI models for conversation and for matching questions to experts, and on the network side for someone to find experts in subject areas the platform doesn’t yet cover.

Forerunner’s Green said that SuperMe is well positioned to build a platform where AI collaborates with humans.

“I think it’s one of the reasons why this is very timely, and it’s good to have something like this establishing itself when we’re relatively early in this journey,” she told TechCrunch over a call. “Because there’s going to be so many lessons that are going to be learned, and they need to be shared so that people can come along in a productive way. I think that the balance here is to build a network so that you do have a true sharing component where there are incentives and benefits for everybody to contribute what they know.”

Here’s another early-stage, agent-management startup to watch because of the backgrounds of the founders and the sheer number of big names it has already attracted as investors: Mintive.

Mintive has raised $11 million total, including a newly announced $7 million seed led by Cowboy Ventures. Additional investors include Coatue, Operator Collective, and Hustle Fund; angels include Okta co-founder Frederic Kerrest, Discovery Loop co-founder and former Google Chief Scientist Jeff Dean, and OpenAI co-founder Andrej Karpathy, among others.

Mintive was founded by Jiquan Ngiam (pictured above, far right) and Vijay Vasudevan (pictured above far left). Ngiam, Mintive’s CEO, was the founding engineer of Coursera (employee No. 1, hired by Andrew Ng). Ngiam later worked on neural networks at Google Brain. Vasudevan, the company’s CTO, is best known as one of the co-creators of TensorFlow while he was at Google Brain. That was the OG AI training platform that predated PyTorch.

Mintive is joining a crowded field of products that promise to govern, provision, and offer identity management to enterprise AI agents. Its platform includes an MCP gateway that allows IT professionals to authorize what the cloud services agents may access and what actions they can perform on them. It also offers an LLM gateway that routes to different models and can cap costs; handles agent identities and authorizations that may differ from the employees that they serve; and an agent monitor that tracks all agents across the company. The company also allows enterprises to create agents that proactively perform long-running tasks.

Still, it’s hard to imagine a field with more competition. Agent management systems are offered by Amazon, Microsoft, OpenAI, Anthropic, plus as point products galore (from companies as diverse as LangChain, New Relic, and Rippling, to name a few). Then there are countless startups working on the idea (40+ from Y Combinator alone).

So, what’s Mintive’s secret sauce? Given the founders’ long history in building AI for enterprises, they have designed their system to be especially easy to use, Ngiam says. Set-up amounts to clicking a few buttons, the startup promises.

“I think what our customers look at and they are a bit shocked at, is how quick it is and easy to deploy,” Ngiam tells TechCrunch. “We have really thought through all the ways this can be done, all the access control, permissioning layer, to make it so easy for them to deploy it, that all the complexities are hidden away. You don’t think about sandboxes, you don’t think about harnesses, you don’t think about memory because we manage that for you.”

That pitch has won over 50 enterprise customers so far, Mintive says, since the startup launched its product in February. These include Ro, Modern Treasury, the University of Michigan, Coursera, Apax, and Stability AI.

But it was the founders’ industry experience that won over Cowboy Ventures managing partner Aileen Lee.

“AI is going to be able to power incredible things, but we need more infrastructure that’s going to give us governance and visibility and control,” Lee told TechCrunch.

Lee also says she particularly respects that Ngiam has been careful by only raising $11 million. This compares to other agentic-building startups raising vast amounts at unicorn valuations. As just one example: Prime Intellect raised $130 million for its Series A at $1 billion valuation.

“I feel really lucky to have won the chance to work with him,” said Lee. “He is a grown-up and so he was not a person chasing a ‘vanity’ round.”

Lee added: “I think a lot of companies are being set up for failure by chasing round size, rather than, ‘how much do I need to actually really move the needle for the company?’ And so, yes, in a time when people are raising unprecedented amounts of money, we’re just kind of doing it the brick by brick, old-fashioned way.”

The race to build the winning AI assistant is on, and investors are spreading out their bets. On Thursday, another AI app, Wabi, now describing itself as a “personal agent,” said it has raised $55 million at a $1.25 billion valuation.

Founded by Eugenia Kuyda, who previously founded the AI companion startup Replika, Wabi originally focused on allowing anyone to build apps using prompts. Last month, however, the startup announced a pivot of sorts to become an AI messenger.

The idea now is that the new Wabi — or Wabi 2.0, as it was called — could build the interfaces you need to get things done as part of its role as your personal AI agent.

The company suggests Wabi could be used to build mini apps for things like meal planning, tracking local restaurants, planning date nights, creating fitness plans, designing specialized trivia games, and more.

Image Credits:Wabi

The additional funding indicates that Wabi’s investors agree that its new agentic angle has some potential, especially given the market around AI agents is hot right now, with viral AI agent Instinct pulling in another billion at a $10 billion valuation in September.

Even Tolan, which makes a cute, Muppet-like AI companion to help people navigate life, just raised $30 million.

“Eugenia has a unique sense for where AI is headed and a track record of great taste in building it,” said Alex Lim, general partner at IVP, which led the new round, in a statement. “She has a clear vision for how people will interact with AI, one that goes far beyond the text box or command line. Wabi is the expression of that vision, and we’re thrilled to support her.”

Other participants in the round included Andreessen Horowitz (a16z), Forerunner, Amazon Alexa Fund, and angel investors like Shopify CEO Tobi Lütke, Salesforce CEO Marc Benioff, former Yahoo! head Marissa Mayer, former Twitter CEO Dick Costolo, and one half of The Chainsmokers, Alex Pall.