In Brief

Prediction market Kalshi is making it easier for its users to have conversations on Meta’s social network Threads. Kalshi now offers a share option that will automatically embed the relevant prediction market chart into a Threads post.

Whether people want to discuss who’s going to win Best Picture or which reality TV contestant is going to go home (and possibly bet on the outcome on Kalshi), “with this integration, people can share their opinions alongside the forecasts they’re seeing on Kalshi,” the company said in a blog post.

It’s a move that echoes a successful social media strategy for both Kalshi and its biggest rival, Polymarket, on X. However, things have gotten complicated for Kalshi on X recently. In June, X named Polymarket as its “official” prediction market partner.

Last month, Kalshi removed its affiliate badges from X accounts run by its sponsored traders. This came after X enacted a policy that prohibits sponsored accounts from posting about sports betting. That policy was adopted after the prediction markets were reportedly busted for partnering with fake sports insiders who spread misinformation.

While the share button is not as significant as what the prediction markets have with X, it is a vote of confidence in the X rival just a couple of months after user data appeared to show Threads growing faster than X.

Amazon announced on Tuesday that it’s expanding access to its healthcare AI assistant to its website and app. The assistant, called Health AI, was previously only available on the app for One Medical, the healthcare company Amazon acquired for $3.9 billion in 2023.

Health AI can answer questions, explain health records, manage prescription renewals, book appointments, and more. Users don’t need to be Prime subscribers or One Medical members to use the assistant, Amazon says.

Amazon says that while Health AI can answer general health questions without having access to an individual’s medical information, it’s designed to serve as a personalized health assistant that can offer more tailored guidance and take actions, such as connecting you with healthcare professionals and treatments.

Of course, there are risks associated with sharing your health information with AI, and researchers are warning against doing so, cautioning that companies are pulling user conversations for training.

Amazon said in its announcement that it trains “Health AI models on abstracted patterns without directly identifying information.” For example, if multiple patients ask about medication interactions, Amazon says it might use these patterns, while keeping patient names private, to improve how Health AI responds to similar questions.

The company also said that all interactions with Health AI happen within a HIPAA-compliant environment, and that conversations are protected by “encryption and strict access controls.” However, Amazon didn’t share exactly how conversations are encrypted and who has access to them. TechCrunch has reached out for specifics.

With a user’s permission, Health AI gets access to their health information through the Health Information Exchange, which is the nationwide secure system for sharing patient medical data.

Techcrunch event Disrupt 2026: The tech ecosystem, all in one room Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to $400. Save up to $300 or 30% to TechCrunch Founder Summit 1,000+ founders and investors come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Offer ends March 13. San Francisco, CA | REGISTER NOW

Health AI can then interpret your lab results, diagnoses, and medical records to provide accurate, personalized answers about your symptoms and medication, Amazon says.

Image Credits:Amazon

If you need professional care, Health AI can connect you to a One Medical provider. Prime members in the U.S. using Health AI receive up to five free direct-message care consultations with a One Medical provider for over 30 common conditions, including cold and flu, allergies, acid reflux, pink eye, UTIs, erectile dysfunction, anti-aging skin care, hair loss, and more. Non-Prime members can connect with One Medical providers through Amazon’s pay-per-visit option.

Users can sign up for Health AI on the Amazon Health page. As Amazon expands access, users will receive an email once they are able to access the assistant.

Once users have access, they need to create a or sign in to their personal Amazon Health profile. They can then start a conversation by typing their health question to Health AI on Amazon.com or in the Amazon app.

Users can ask questions like “Can you explain my recent cholesterol results and what they mean for me?” or “I’m feeling congested and have a sore throat. What should I do?”

Amazon’s expansion of Health AI comes as popular AI services have quickly moved into the healthcare space. In January, OpenAI released ChatGPT Health, a version of its chatbot tailored to answer health questions. A week later, Anthropic announced its own healthcare-focused product, Claude for Healthcare.

A former employee of Elon Musk’s Department of Government Efficiency reportedly stole Americans’ personal data from the U.S. Social Security Administration and stored it on a thumb drive, according to a whistleblower complaint reported by The Washington Post.

The former DOGE software engineer told co-workers at his new job that he “possessed two tightly restricted databases of U.S. citizens’ information” and was planning to use the information at his new company, according to the report, which added that the Social Security Administration’s inspector general is investigating the whistleblower complaint.

The former DOGE employee, whom The Washington Post did not name, worked at the Social Security Administration last year. In October he then left to work at a government contractor, where he told colleagues that he had obtained two databases, called “Numident” and the “Master Death File,” according to The Post, which reported that the databases could include records for “more than 500 million living and dead Americans, including Social Security numbers, places and dates of birth, citizenship, race and ethnicity, and parents’ names.”

The man also reportedly claimed that he previously had unrestricted “God-level” access to the SSA’s systems.

A spokesperson for the Social Security Administration, which is still under control by DOGE, denied that a former employee stole data on U.S. citizens. The spokesperson said the Washington Post was “desperate for clicks and eager to publish fake news to scare seniors.” The inspector general’s office; which is independent from the Trump administration, did not immediately respond to a request for comment.

This is the latest case of a suspected breach of Americans’ personal data linked to DOGE and its incursion into the Social Security Administration. In January, two DOGE members were suspected of accessing and sharing Social Security numbers that were off-limits to them as part of an effort to aid an advocacy group that intended to “overturn election results in certain States,” according to a lawsuit.

Another whistleblower within the agency last year said that DOGE members put Americans at risk by uploading hundreds of millions of Social Security records to a vulnerable cloud server. Also last year, a judge blocked DOGE from accessing SSA systems, accusing the Musk-led agency of being “essentially engaged in a fishing expedition” in search of fraud.

Techcrunch event Disrupt 2026: The tech ecosystem, all in one room Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to $400. Save up to $300 or 30% to TechCrunch Founder Summit 1,000+ founders and investors come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Offer ends March 13. San Francisco, CA | REGISTER NOW

Several DOGE members were installed at the Social Security Administration after Trump took office last year. According to The Washington Post, at least a dozen DOGE employees, most of them technical staff or engineers, worked at the agency, though their actual roles and activities were not communicated to the rest of the staff.

On Friday night, my boyfriend and I sat on the couch for a refreshing evening of doing nothing together. We tuned into a baseball game, he picked up my guitar, and I eagerly booted up “Pokémon Pokopia,” the 30-year-old franchise’s new cozy life simulator game, which is unlike anything we’ve seen from Pokémon before.

I narrated my experience as I played, explaining the process of constructing habitats to increase the comfort levels of my Pokémon friends, a primary objective of the game.

“Onix is stuck in a cave, but I can’t break through the walls, so Squirtle suggested throwing a party to make it rain to soften the rocks,” I told my boyfriend as I played. “But Squirtle and I don’t know what ‘celebration’ means, so we have to ask Professor Tangrowth what it means to ‘party.’”

I rejoiced when I finally made it rain and awakened Kyogre — but then Charmander, who calls me “bestie,” discovered that the rain makes the flame on its tail go out, so I had to build a little hut for shelter with the help of our pals Timburr and Hitmonchan.

Suddenly, it was 11:30 p.m. I only looked up because the baseball game was about to end. To my horror, my boyfriend had fallen asleep on the couch beside me.

I did not realize he was asleep. I was so engrossed in building habitats for my Pokémon pals that I didn’t notice that he had stopped responding to my commentary … since he was no longer awake. While he drifted in and out of a light couch snooze, I had never stopped relaying a detailed play-by-play of how I was restoring a seaside habitat for Magikarp. I was completely oblivious.

I was, and am, embarrassed that this happened. For my own good, I have no choice but to believe that I committed this faux pas not because I am an inattentive partner, but because “Pokopia” is simply too good a game, and thus, it is not my fault that I paid more attention to the fictional Onix stuck in a cave than the actual human being beside me. (You should’ve seen how helpless that Onix looked! How long had he been stuck in there?)

“Pokopia” is like an “Animal Crossing,” “Stardew Valley,” and “Minecraft” hybrid, but set in Pokémon’s Kanto region, which has now become an apocalyptic wasteland. Given the bleak setting, it’s impressive that “Pokopia” is still firmly in the category of cozy gaming.

I’m not alone in my obsession with “Pokopia.” The game seems to be so popular that it surpassed sales expectations, leading Amazon to bump the cost of physical game copies by $10, bringing it to a whopping $80 (the game is also available as a digital download). It’s also the first Switch 2 exclusive game that is generating enough buzz to make people go out and upgrade to the new console.

The last few main series Pokémon games, like “Pokémon Scarlet” and “Pokémon Violet,” were met with lukewarm reception — the games were buggy, and the open-world layout wasn’t quite intriguing enough to compensate for how rushed they felt. Even as a lifelong Pokémon fan who will dutifully buy any game the franchise puts out, I’ve found the recent installments to be fun, but they lose my attention once I complete the main storyline. Yet “Pokopia” has far exceeded my wildest expectations with how expansive and thoughtfully designed it is.

There are four main regions in “Pokopia,” plus a sandbox version of Palette Town for group play. If I had to guess, I’d say I’ve played a solid 20 hours of “Pokopia” since it came out less than a week ago (whoops!), and I am less than halfway through the main story. It feels gloriously endless, even if it’s not — but even then, I could definitely see the developers releasing additional regions to explore as part of a DLC pack, which I would gladly pay for despite the game’s already high price of $70.

Few games have put me in a flow state like this. It’s hard not to compare the feeling to when “Animal Crossing: New Horizons” first came out, but this time, we thankfully are not experiencing the onset of a pandemic lockdown that would change our lives indefinitely.

A lot in the world has improved since “Animal Crossing” came out — yay, coronavirus vaccines! — and yet, so much feels the same. Donald Trump is president again. The federal government is siccing armed agents on ordinary people rallying for civil rights. Extreme weather is becoming the norm. Things still feel bad.

Like “Animal Crossing,” playing “Pokopia” is an escape and distraction, yet it’s grounded in our actual world in a way that your island getaway with Tom Nook is not.

In the post-apocalyptic Kanto region of “Pokopia,” you play as a Ditto who has transformed to look like its former trainer, who is inexplicably missing — in fact, all of the humans are gone, and when you randomly appear in a cave with Professor Tangrowth, the graying vine Pokémon hasn’t seen another creature in many years.

It’s not immediately clear what happened that made Kanto evolve into a barren wasteland, but as your Ditto explores the ruins and restores habitats to find new Pokémon, you encounter scraps of diary entries, newspaper articles, and letters that help you piece together what happened: There was some sort of disastrous climate event, and as a result, all of the humans are gone. Pikachu appears in the game as “Peakychu,” a pale creature who lost its ability to produce electricity, and Snorlax has been solitarily sleeping in a cave long enough that it’s become part of the landscape, covered in moss. Yikes.

Image Credits:Pokémon Pokopia

The apocalyptic mystery makes each new morsel of information feel more exciting, if not foreboding.

“We all know that everyone’s beloved music streaming services are being forced to shut down one by one due to the steep rise of server fees all over the world,” one note from an old Poké Mart says. “While music lovers are still mourning the loss of these streaming services, it isn’t bad news in the world of music!”

The note continues to explain the return to CDs that “our great-great-grandparents” used, which don’t charge a subscription fee, no matter how many times you listen to them.

It’s funny that Nintendo is poking fun at the broken model of music streaming, but the bit about server costs feels a bit too real for this moment. Since fast-growing AI tools require so much computing power to operate, there are nearly 3,000 energy-intensive data centers under construction in the U.S., which will add to the 4,000 already in operation. The demand for more computing power is so high that the tech industry is facing a RAM shortage severe enough to bump the price of new MacBook Pros by up to $400.

Climate crisis? Server costs? Broken music streaming models? It almost feels as though Nintendo is trying to say something about the current state of the world.

But while “Animal Crossing” is pure escapism, “Pokopia” at least gives you the sensation of actually rehabilitating a broken world. It’s unsettling to see Vermilion City in ruins — but that only makes it more rewarding when you work with the other Pokémon to rediscover electricity and illuminate the landscape, eclipsing the dark clouds with a burst of light.

YouTube has hit some impressive numbers in 2025.

According to new estimates from research firm Moffett Nathanson — as reported by The Hollywood Reporter — the platform pulled in a staggering $40.4 billion in ad revenue, which is more than Disney, NBC, Paramount, and Warner Bros. Discovery’s (WBD’s) combined ad revenue, whose total came to $37.8 billion.

This shift is a clear sign of just how much content consumption has changed over the years and how advertisers have adapted their strategies.

YouTube’s ad revenue for 2025 not only topped the combined haul from the four major Hollywood studios, but also marked a huge turnaround from the previous year. In 2024, YouTube’s ad revenue of $36.1 billion fell short of the $41.8 billion in ad revenue collectively earned by Disney, NBCU, Paramount, and WBD. The tables have now turned.

For decades, these studios have been the kings of entertainment, captivating audiences with big-budget films and hit TV shows. But as traditional studios struggle with shrinking linear TV audiences and the ever-rising costs of production, YouTube is speeding ahead. Even as these companies pour millions into their own streaming platforms, it’s getting harder to keep up with YouTube’s momentum.

Last month, parent company Alphabet reported that YouTube’s total revenue in 2025 soared to $60 billion. A big portion of YouTube’s revenue now comes from subscriptions, which include services like YouTube TV, YouTube Premium, YouTube Music, and NFL Sunday Ticket. (For context, that’s well above Netflix’s reported $45.2 billion for the full year.)

Disney, NBC, Paramount, and WBD are also heavily dependent on subscriptions; for instance, Disney’s media business pulled in a total of $60.9 billion in revenue last year, when you include subscriptions.

Techcrunch event Disrupt 2026: The tech ecosystem, all in one room Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to $400. Save up to $300 or 30% to TechCrunch Founder Summit 1,000+ founders and investors come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Offer ends March 13. San Francisco, CA | REGISTER NOW

That said, YouTube’s ad revenue still trails behind tech giants like Meta, which pulled in $196.2 billion in ad revenue in 2025. Still, advertisers are flocking to YouTube because that’s where audiences (especially younger viewers) are spending their time. YouTube’s ad revenue reached $11.4 billion in the fourth quarter.

YouTube is also stepping up its investment in AI. This week, the company announced that it’s expanding its likeness-detection technology to a pilot group of government officials, politicians, and journalists. The tech identifies AI-generated deepfakes and lets users request the removal if they believe it violates YouTube policy.

With the top app stores flooded with AI apps, developers may think the best bet for turning a profit is to integrate artificial intelligence technology into their own products. However, a new study focused on the subscription app ecosystem across iOS, Android, and web is calling that assumption into question.

RevenueCat, a company that offers subscription management tools used by over 75,000 app developers, said in its 2026 State of Subscription Apps Report that AI integration is not a guarantee of long-term retention. Instead, AI-powered apps struggle to retain subscribers, with people canceling their annual subscriptions — a metric known as churn — 30% faster than non-AI apps, at the median, according to the report.

The report is based on an analysis of the subscription app providers that use RevenueCat’s tools to manage their more than 1 billion in-app transactions, generating more than $11 billion in revenue for developers annually. As one of the more popular tools in this space, its data represents a healthy sample in terms of trend analysis.

Among the many interesting findings, the report noted that most of the apps using the company’s platform are not yet powered by AI. AI-powered apps account for 27.1% of apps across all categories, compared with 72.9% for non-AI apps. Still, it’s a growing category, as roughly one in four apps is now AI-powered.

(To be clear, the AI-powered apps category includes the popular AI chatbots, like ChatGPT and Gemini, as well as any app that markets itself as being AI-powered.)

REvenuecat: AI vs. Non-AI apps by category. Image Credits:RevenueCat

Photo & Video apps have the biggest share (61.4%) of AI-powered apps, while gaming has the smallest share at 6.2%. Travel (12.3%) and Business (19.1%) are also low-AI segments.

The more surprising figures are around AI apps’ ability to retain their paying customers. AI apps underperform on retention at both a monthly and annual level, RevenueCat’s data shows.

Annual retention, a metric focused on the app’s ability to retain subscribers after 12 months, was 21.1% for AI apps, compared with a higher 30.7% for non-AI apps. Monthly, AI apps saw 6.1% retention rates versus 9.5% for non-AIs — a difference of 3.4 percentage points.

The only area where AI led on retention was on the weekly front, where AI apps had 2.5% retention rates compared with 1.7% for non-AI apps. It’s worth noting that weekly subscriptions are not the most popular option for AI apps.

Image Credits:RevenueCat

These metrics could be influenced by the rapidly changing state of AI technology, which could see users hopping between different AI apps more quickly, as they try to find the one that has the most current technology under the hood.

AI vs. non-AI apps by subscription plan type. Image Credits:RevenueCat

As customers experiment with a growing number of AI apps, they’re also more likely to find that some don’t meet their needs. The report notes that AI apps have 20% higher refund rates (4.2% vs. 3.5% at the median) than non-AI apps do.

The upper bound of refund rates for AI apps is also higher (15.6% vs. 12.5%), suggesting there’s “greater volatility in realized revenue and deeper issues in user value, experience, and long-term quality,” the report notes.

Image Credits:RevenueCat

There are some benefits to being in the AI-powered apps cohort, the data indicates.

RevenueCat found that AI apps convert users from trials to paid customers 52% better than non-AI apps (8.5% vs. 5.6% at the median), and AI apps monetize their downloads around 20% better than non-AI apps (2.4% to 2% at the median).

AI apps also generate 39% or higher monthly realized lifetime value (RLTV), a metric that measures the actual net value of an average paying user over time. AI apps’ median on this metric is $18.92 per month, compared with $13.59 for non-AI apps. AI apps also sustain a 41% or higher RLTV on an annual basis, at $30.16 vs. $21.37, also at the median.

The overall takeaway from the report’s findings is that AI can drive strong, early monetization, but these apps are struggling to sustain their value with customers over time.

Google, Tesla, and data center developer Verrus are among a group of companies arguing that the electrical grid is being underutilized and they want everyone — especially politicians — to know about it.

The three companies along with HVAC giant Carrier, virtual power plant company Renew Home, distributed energy resource developer Sparkfund, and smart electrical panel startup Span founded a new group called Utilize to get that message across. The group, which launched Tuesday, is advocating to change the way the grid is built and used. The group points out, correctly, that the grid is designed for brief bursts of high demand; most of the time there’s lots of capacity that goes unused.

Utilize thinks that should change. The group argues that smarter ways to use that capacity already exist. Utilize name-checks a number of those solutions, including battery storage, demand response, and virtual power plants, all of which have emerged en masse over the last decade, but remain under utilized. (Oh, that’s where the name comes from.)

In many cases, those new technologies are used to improve the grid’s resilience. Take Texas’ grid, for example, which has fared better during recent cold snaps following an increase in battery storage in the state. Still, many regulators and politicians remain wary of these new technologies, opting instead to stick with familiar options like centralized fossil fuel power plants.

Utilize says it will “advocate for policies” that encourage more widespread adoption of the new technologies, which also benefit those involved.

Each member occupies a niche piece of the grid. On the sell side, Tesla sells batteries and solar panels, Span sells an electrical panel that can react to changing loads, Carrier makes heat pumps, and Sparkfund and Renew Home build and aggregate distributed energy resources. On the buy side, Google and Verrus have enormous power needs to keep their servers humming.

The organization calls itself a “coalition,” which is pretty squishy language. Utilize already touts one legislative win, saying that “some members of Utilize” backed a bill in Virginia that would require utilities to quantify and disclose how the grid is being used.

Techcrunch event Disrupt 2026: The tech ecosystem, all in one room Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to $400. Save up to $300 or 30% to TechCrunch Founder Summit 1,000+ founders and investors come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Offer ends March 13. San Francisco, CA | REGISTER NOW

That wording suggests that while Utilize might be pushing for policy changes, it probably isn’t lobbying directly, at least yet. TechCrunch did not receive a reply to inquiries sent to Utilize and the Commonwealth of Virginia regarding the organization’s status as a lobbyist.

Advocacy organizations are no stranger to the utility industry, but the combination of new technology paired with buy- and sell-side companies makes Utilize something different. Changing the way the grid is regulated is a long game, but if they don’t start now, it’ll be too late.

On Tuesday, OpenAI introduced dynamic visual explanations, a new ChatGPT feature that allows users to see how formulas, variables, and mathematical relationships change in real time.

Instead of just reading an explanation or looking at a static diagram, users can now engage directly with interactive visuals. For example, when exploring the Pythagorean theorem, you can adjust the lengths of the triangle’s sides and instantly watch the hypotenuse update. Users can adjust numbers and variables and see changes reflected immediately.

To try out the feature, ask ChatGPT questions like “What is a lens equation?” or “How can I find the area of a circle?” ChatGPT will respond not only with an explanation but also with an interactive module you can manipulate.

Currently, these visuals are available for over 70 math and science topics, including the binomial square, Charles’ law, area of a circle, compound interest, Coulomb’s law, difference of squares, exponential decay, Hooke’s law, kinetic energy, linear equations, and Ohm’s law.

OpenAI plans to expand the feature later on to include more interactive topics. It’s available to all logged-in ChatGPT users.

Image Credits:OpenAI

The launch of dynamic visual explanations is notable in that it shifts the tool’s role somewhat — rather than just delivering answers, it asks users to engage with the underlying concepts directly. Whether that translates to deeper understanding will likely depend on how it’s used.

As AI continues to transform the way people approach learning, it has sparked heated debates in the education community. Though some educators worry about overreliance on AI, many teachers and students have already embraced these tools, integrating them into daily study routines. According to OpenAI, more than 140 million people use ChatGPT each week for help with math and science, subjects that have traditionally challenged learners.

Techcrunch event Disrupt 2026: The tech ecosystem, all in one room Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to $400. Save up to $300 or 30% to TechCrunch Founder Summit 1,000+ founders and investors come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Offer ends March 13. San Francisco, CA | REGISTER NOW

Other major AI companies are also exploring similar interactive features. In November, Google’s Gemini launched its own interactive diagrams and visuals.

This new feature joins other recent ChatGPT educational tools, such as study mode, which guides users through problems step by step. There’s also QuizGPT, where users can make flashcards and ask ChatGPT to quiz them on any topic before exams.

A mass hacking campaign targeting iPhone users in Ukraine and China used tools that were likely designed by U.S. military contractor L3Harris, TechCrunch has learned. The tools, which were intended for Western spies, wound up in the hands of various hacking groups, including Russian government spooks and Chinese cybercriminals.

Last week, Google revealed that over the course of 2025, it discovered that a sophisticated iPhone-hacking toolkit had been used in a series of global attacks. The toolkit, dubbed “Coruna” by its original developer, was made of 23 different components first used “in highly targeted operations” by an unnamed government customer of an unspecified “surveillance vendor.” It was then used by Russian government spies against a limited number of Ukrainians and finally by Chinese cybercriminals “in broad-scale” campaigns with the goal of stealing money and cryptocurrency.

Researchers at mobile cybersecurity company iVerify, which independently analyzed Coruna, said they believed it may have been originally built by a company that sold it to the U.S. government.

Two former employees of government contractor L3Harris told TechCrunch that Coruna was, at least in part, developed by the company’s hacking and surveillance tech division, Trenchant. The two former employees both had knowledge of the company’s iPhone hacking tools. Both spoke on condition of anonymity because they weren’t authorized to talk about their work for the company.

“Coruna was definitely an internal name of a component,” said one former L3Harris employee, who was familiar with iPhone hacking tools as part of their work at Trenchant.

“Looking at the technical details,” this person said, referring to some of the evidence Google published, “so many are familiar.”

Contact Us Do you have more information about Coruna, or other government hacking and spyware tools? From a non-work device, you can contact Lorenzo Franceschi-Bicchierai securely on Signal at +1 917 257 1382, or via Telegram, Keybase and Wire @lorenzofb, or Do you have more information about Coruna, or other government hacking and spyware tools? From a non-work device, you can contact Lorenzo Franceschi-Bicchierai securely on Signal at +1 917 257 1382, or via Telegram, Keybase and Wire @lorenzofb, or by email .

The former employee said the overarching Trenchant toolkit housed several different components, including Coruna and related exploits. Another former employee confirmed that some of the details included in the published hacking toolkit came from Trenchant.

L3Harris sells Trenchant’s hacking and surveillance tools exclusively to the U.S. government and its allies in the so-called Five Eyes intelligence alliance, which includes Australia, Canada, New Zealand, and the United Kingdom. Given Trenchant’s limited number of customers, it’s possible that Coruna was originally acquired and used by one of these governments’ intelligence agencies before falling into unintended hands, though it’s unclear how much of the published Coruna hacking toolkit were developed by L3Harris Trenchant.

An L3Harris spokesperson did not respond to a request for comment.

A globe-trotting iPhone-hacking toolkit

How Coruna went from the hands of a Five Eyes government contractor to a Russian government hacking group and then to a Chinese cybercrime gang is unclear.

But some of the circumstances appear similar to the case of Peter Williams, a former general manager at Trenchant. From 2022 until he resigned in mid-2025, Williams sold eight company hacking tools to Operation Zero, a Russian company that offers millions of dollars in exchange for zero-day exploits, meaning vulnerabilities that are unknown to the affected vendor.

Williams, a 39-year-old Australian citizen, was sentenced to seven years in prison last month, after he admitted to stealing and selling the eight Trenchant hacking tools to Operation Zero for $1.3 million.

The U.S. government said Williams, who took advantage of having “full access” to Trenchant’s networks, “betrayed” the United States and its allies. Prosecutors accused him of leaking tools that could have allowed whoever used them to “potentially access millions of computers and devices around the world,” suggesting the tools relied on vulnerabilities affecting widely used software like iOS.

Operation Zero, which was sanctioned by the U.S. government last month, claims to work exclusively with the Russian government and local companies. The U.S Treasury claimed that the Russian broker sold Williams’ “stolen tools to at least one unauthorized user.”

That would explain how the Russian espionage group, which Google has only identified as UNC6353, acquired Coruna and deployed it on compromised Ukrainian websites so that it would hack certain iPhone users from a specific geolocation who unwittingly visited the malicious site.

It is possible that once Operation Zero acquired Coruna and potentially sold it to the Russian government, the broker then resold the toolkit to someone else, perhaps another broker, another country, or even directly to cybercriminals. The Treasury alleged that a member of the Trickbot ransomware gang worked with Operation Zero, tying the broker to financially motivated hackers.

At that point, Coruna may have passed to other hands until it reached Chinese hackers. According to U.S. prosecutors, Williams recognized code that he wrote and sold to Operation Zero later being used by a South Korean broker.

the logo Kaspersky made for Operation Triangulation next to the L3Harris logo. Image Credits:Kaspersky and L3Harris

Operation Triangulation

Google researchers wrote on Tuesday that two specific Coruna exploits and underlying vulnerabilities, called Photon and Gallium by their original developers, were used as zero-days in Operation Triangulation, a sophisticated hacking campaign allegedly used against Russian iPhone users. Operation Triangulation was first revealed by Kaspersky in 2023.

Rocky Cole, the co-founder of iVerify, told TechCrunch that “the best explanation based on what’s known right now” points to Trenchant and the U.S. government being the original developers and customers of Coruna. Although, Cole added, he isn’t claiming this “definitively.”

That assessment, he said, is based on three factors. The timeline of Coruna’s use lines up with Williams’ leaks; the structure of three modules — Plasma, Photon, and Gallium — found in Coruna bear strong similarities with Triangulation; and Coruna reused some of the same exploits used in that operation.

According to Cole, “people close to the defense community” claim Plasma was used in Operation Triangulation, “although there’s no public evidence of that.” (Cole previously worked at the U.S. National Security Agency.)

According to Google and iVerify, Coruna was designed to hack iPhone models running iOS 13 through 17.2.1, released between September 2019 and December 2023. Those dates line up with the timeline of some of Williams’ leaks and the discovery of Operation Triangulation.

One of the former Trenchant employees told TechCrunch that when Triangulation was first revealed in 2023, other employees at the company believed that at least one of the zero-days caught by Kaspersky “were from us, and potentially ‘ripped out’ of” the overarching project that included Coruna.

Another breadcrumb that points to Trenchant — as security researcher Costin Raiu noted — is the use of bird names for some of the 23 tools, such as Cassowary, Terrorbird, Bluebird, Jacurutu, and Sparrow. In 2021, The Washington Post revealed that Azimuth, one of the two startups later acquired by L3Harris and merged into Trenchant, had sold a hacking tool called Condor to the FBI in the infamous San Bernardino iPhone cracking case.

After Kaspersky published its research on Operation Triangulation, Russia’s Federal Security Service (FSB) accused the NSA of hacking “thousands” of iPhones in Russia, targeting diplomats in particular. A Kaspersky spokesperson said at the time that the company did not have information on the FSB’s claims. The spokesperson did note that “indicators of compromise” — meaning evidence of a hack — identified by the Russian National Coordination Centre for Computer Incidents (NCCCI) were the same ones that Kaspersky had identified.

Boris Larin, a security researcher at Kaspersky, told TechCrunch in an email that “despite our extensive research, we are unable to attribute Operation Triangulation to any known [Advanced Persistent Threat] group or exploit development company.”

Larin explained that Google linked Coruna to Operation Triangulation because they both exploit the same two vulnerabilities — Photon and Gallium.

“Attribution cannot be based solely on the fact of exploitation of these vulnerabilities. All the details of both vulnerabilities have long been publicly available,” and thus anyone could have taken advantage of them, he said, adding that those two shared vulnerabilities “are just the tip of the iceberg.”

Kaspersky never publicly accused the U.S. government of being behind Operation Triangulation. Curiously, the logo that the company created for the campaign — an apple logo composed of several triangles — is reminiscent of the L3Harris logo, and Trenchant’s own logo is made of two triangles. It may not be a coincidence. Kaspersky has previously said it wouldn’t attribute a hacking campaign publicly while quietly signaling that it actually knew who was behind it, or who provided the tools for it.

In 2014, Kaspersky announced that it had caught a sophisticated and elusive government hacking group known as “Careto” (Spanish for “the Mask”). The company only said the hackers spoke Spanish. But the illustration of a mask that the company used in its report included the red and yellow colors of Spain’s flag, bull’s horns and nose ring, and castanets.

As TechCrunch revealed last year, Kaspersky researchers had privately concluded that “there was no doubt,” as one of them put it, that Careto was run by the Spanish government.

On Wednesday, cybersecurity journalist Patrick Gray said on an episode of his podcast Risky Business that he thought — based on “bits and pieces” he was confident about — that what Williams leaked to Operation Zero was the hacking kit used in the Triangulation campaign.

Apple, Google, and Operation Zero did not respond to requests for comment.

This post originally published at March 9 at 6:56 p.m. PT

Just a couple of years ago, AI agents were mostly chatbots that could use basic tools. People were curious, but given concerns around reliability and security, as well as cost, the tech remained in the realm of early adopters.

How things have changed. Coding agents like Claude Code, Codex, and Cursor initially saw the most traction, spurring adoption among programmers around the world, but today we have people using AI agents to do everything from debugging at scale and building marketing campaigns, to managing calendars and scheduling meetings. OpenClaw‘s blockbuster debut earlier this year only sped things up, opening up access to AI agents by letting users run their own localized and personalized agent round the clock.

And if the tech industry is to be believed, AI agents are set to become as numerous as real people on the internet, using software and services, talking and shopping on your behalf, and generally automating a wide swathe of work.

AgentMail, a startup out of San Francisco, sees that future playing out for certain, which is why it has built an email service designed specifically for AI agents. The company provides an API platform that lets you give AI agents their own email inboxes, with support for two-way conversations, parsing, threading, labeling, searching, and replying.

The company on Tuesday said it had raised $6 million in a seed funding round led by General Catalyst, with participation from Y Combinator, Phosphor Capital, and angel investors Paul Graham, Dharmesh Shah (CTO of HubSpot), Paul Copplestone (CEO of Supabase), and Karim Atiyeh (CTO of Ramp).

Alongside the funding, AgentMail also announced an onboarding API that you can point your AI agent to so it can directly sign up and create an email inbox for itself. The platform also lets you set up and manage inboxes, permissions, allowlists, and API keys manually.

According to co-founder and CEO Haakam Aujla (pictured above, far right), AgentMail was built from the ground up to provide AI agents a similar inbox experience as people get with services like Gmail or Outlook — except without the UI elements humans need. (Note: The platform provides a perfectly human-usable interface, too, for managing the various agent inboxes, and reading and sending emails.)

“When you open Gmail, you have a bunch of threads, and inside each thread, you can have many messages; those messages can have attachments. You want to be able to label them, search them, filter them, reply, forward,” Aujla told TechCrunch. “We thought we wanted our agents to be able to do that, but they shouldn’t have to, you know, click buttons on a screen, because that’s pretty clunky for agents to do. They should just be able to make API calls.”

Since launching as part of Y Combinator’s Summer 2025 batch, the company has attracted tens of thousands of human users, and hundreds of thousands of “agent users,” Aujla said, as well as more than 500 B2B customers.

The initial days were slow, however, as AI agents hadn’t really taken off yet. AgentMail, therefore, focused on B2B use cases for companies that wanted to things like scale their email communications. But when OpenClaw (then known as Clawdbot) burst onto the scene in late January, AgentMail saw its user count triple that week, and quadruple in February as people started looking for a way to give agents an email inbox so they could do more on their own.

The timing was just right, as traditional email providers like Gmail impose rate and volume limits on their email APIs. AgentMail, meanwhile, provides a pretty generous free tier, in addition to paid plans and enterprise subscriptions.

But there’s an obvious issue with giving email inboxes to AI agents: it makes misuse easy. To counteract abuse, Aujla said AgentMail has a few systems: Agent inboxes can only send 10 emails a day unless they’re authenticated by a person; the platform imposes rate limits if it detects unusual levels of high activity from inboxes; it monitors for bounce rates; and it randomly samples new accounts to filter for sensitive keywords.

Aujla says beyond providing a way for bots to send and receive emails, AgentMail’s larger purpose is to serve as an identity layer for AI agents: “We want to give agents the ability to use email in the same way that humans do, right? But the idea is, what humans use email for is not even communication. It’s your identity […] There are several startups that are trying to build new identity protocols for agents, but our thesis is, let’s just use what already works for humans, and what already is so deeply integrated into the entire internet.”

Aujla concluded, “You give an agent an email address, [and] it can now use essentially any software service that already exists.”

In Brief

OpenAI co-founder Mira Murati’s two-year-old AI research lab has signed a sizable deal with semiconductor giant Nvidia.

Murati’s Thinking Machines Lab announced it entered into a multi-year strategic partnership with AI semiconductor giant Nvidia on Tuesday. The size of the deal was not disclosed and includes the AI research lab deploying at least one gigawatt of Nvidia’s Vera Rubin systems, which was released earlier this year, starting in 2027.

Nvidia is also making a strategic investment in Thinking Machines Lab, which has raised more than $2 billion since its February 2025 founding from investors including Andreessen Horowitz, Accel, and Nvidia, among others, including rival chipmaker AMD’s venture arm.

The seed-stage company is valued at more than $12 billion and is working to build AI models that create reproducible results. The company released its first product, an API called Tinker, in October.

TechCrunch reached out to Thinking Machines Lab and Nvidia for more information regarding the specifics surrounding the deal terms and investment. Thinking Machines Lab declined to comment beyond the release.

The partnership also includes a commitment to develop training and serving systems for Nvidia architecture, according to an Nvidia press release.

“Nvidia’s technology is the foundation on which the entire field is built,” Murati said in the deal’s blog post. “This partnership accelerates our capacity to build AI that people can shape and make their own, as it shapes human potential in turn.”

Thinking Machines Lab has seen a number of recent high-profile exits in its young history. The company’s co-founder, Andrew Tulloch, left the startup for a role at Meta in October. Earlier this year, three additional co-founders, Barret Zoph, Luke Metz, and Sam Schoenholz, left to return to OpenAI.

This deal comes as AI companies remain hungry for any compute that they can get. Nvidia CEO Jensen Huang predicted that companies could spend $3 trillion to $4 trillion on AI infrastructure by the end of the decade.

While we don’t know the value of this specific deal, it’s believable. In 2025, rival OpenAI allegedly inked a historic $300 billion compute deal with Oracle.

Updated after publication to correct that Thinking Machines released its product Tinker last fall.

In a slight capitulation to those who don’t want AI infused into their everyday apps, Google said it’s now offering a toggle that will allow users of its Google Photos app to return to the previous and often faster “classic” search experience instead of the newer AI-powered option known as “Ask Photos.”

The Ask Photos feature, launched in the U.S. in 2024, lets users search their photos using natural language queries, including complex requests. The product’s rollout was briefly paused last summer as the company worked to address issues around latency, following user feedback.

Some Google Photos users never warmed up to the AI-powered experience, complaining that Ask Photos still failed to find some of their photos and that searches were less accurate than before.

While Google offered an option to disable the use of Gemini in Google Photos, it was buried in the settings and was often overlooked.

The company said it will offer users an easier and more visible way to switch between the two search experiences. Via a new toggle button on the search screen, users can turn the Ask Photos AI search off and view the classic results instead. Google said it will still lead with whichever results best fit the user’s query, however.

In the announcement, shared by Google Photos lead Shimrit Ben-Yair, the company suggested that the move was driven by users’ complaints about the Ask Photos feature. In a post on X, Ben-Yair wrote, “We’ve heard your feedback that you want more control over the type of results you see when searching in Google Photos.”

The exec also noted that Google had improved the quality of some of the most popular searches, also based on user feedback.

“We know search in Photos is one of the most loved and used features and we’re committed to getting this experience right, so please keep the feedback coming! It helps us build a more magical experience for everyone,” she said.

Kevin Mandia, who founded the cybersecurity startup Mandiant in 2004 and sold it to Google for $5.4 billion in 2022, has launched a new AI-native cybersecurity startup with what the company claims is a record-breaking funding round.

The new outfit, called Armadin, has raised $189.9 million in combined seed and Series A funding led by Accel, with participation from GV, Kleiner Perkins, Menlo Ventures, 8VC, Ballistic Ventures, and the CIA’s venture arm, In-Q-Tel, it said. The company claims the combined total is a record for a security startup at that early a stage, though it isn’t disclosing its valuation.

While other security startups have raised even slightly bigger Series A rounds, we couldn’t find another one that did so out of the gate. In 2019, for example, password-management company 1Password and privacy compliance company OneTrust both raised $200 million in Series A funding. But 1Password was already 14 years old at the time and OneTrust was three years old and already in growth mode.

Prior to Armadin, Mandia, an internationally recognized security expert, had been a VC at Ballistic Ventures. That’s the security specialist fund co-founded by famed security VC Ted Schlein, formerly of Kleiner Perkins.

Mandia founded Armadin to create autonomous cybersecurity agents, software designed to learn and respond to threats without a human in the middle. He told CNBC that he believes autonomous AI hackers are on the way and that they are to be feared. Security researchers and government agencies have raised similar alarms, warning that AI is already lowering the bar for launching sophisticated attacks.

“When you have AI on offense, what you are going to get is a technology that can think, can learn, can adapt,” he warned, adding that the attackers will be able to complete attacks in minutes that used to take days.

Armadin aims to provide the white hats (aka good-guy security experts) with automated agents so they have their own agentic armies to combat AI-powered attacks run by the black hats (bad guys). Mandia’s co-founders at Armadin are former Google Cloud Security principal engineer Travis Lanham; former Mandiant exec Evan Peña; and former Google SecOps engineer David Slater.

Techcrunch event Disrupt 2026: The tech ecosystem, all in one room Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to $400. Save up to $300 or 30% to TechCrunch Founder Summit 1,000+ founders and investors come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately



Offer ends March 13. San Francisco, CA | REGISTER NOW

TechCrunch has reached out to the company for more information.

Legora, an AI platform for lawyers, is now valued at $5.55 billion following a $550 million Series D set to fuel its growth in the U.S. That’s despite growing competition with rival Harvey, but also with Microsoft Copilot and generalist large language models (LLMs). Publicly listed legal software companies saw their stocks drop when Anthropic unveiled a legal plug-in for Claude.

Legora is built on top of LLMs, and mostly on Claude, but its positioning as a platform that supports lawyers with complex cases gives CEO Max Junestrand some peace of mind. “It’s amazing that everybody can have their own pocket lawyer in Claude, but we’re not solving for the same use case,” he said via livestream at the TechArena conference in Stockholm.

With a focus on embedding itself into its clients’ workflows, Legora’s platform is now used by 800 law firms and legal teams — and investors took note. Its Series D was led by Accel, with participation from existing investors Benchmark, Bessemer, General Catalyst, ICONIQ, Redpoint Ventures, and Y Combinator; and new backers including Alkeon Capital, Bain Capital, Firstmark Capital, Menlo Ventures, Salesforce Ventures, Sands Capital, and Starwood Capital.

There are other signs that investors are bullish about AI legal tech. Legora’s Series D and valuation jump come just a few months after its October 2025 $150 million Series C round led at a $1.8 billion valuation. Its competitor, Harvey, which is backed by a16z, is already valued at $8 billion, and is now reportedly seeking to raise at an $11 billion valuation. According to Dealroom, they are also on almost identical trajectories with regard to revenue.

Both are also branching out globally; Harvey is pushing hard into Europe, and Legora in the opposite direction. Formerly known as Judilica, then Leya, the startup is an alum of Stockholm’s SSE Business Lab, a known breeding ground for unicorns. But after participating in YC’s winter 2024 batch, Legora is now headquartered in New York and keen to keep on pushing in the U.S. market, where its growth exceeded its expectations coming out of Europe.

“It’s nine to one in terms of legal spending; it turns out the Americans love to sue each other much more than we like to do in Europe,” Junestrand joked while speaking to TechArena’s audience. But the team has grown globally — from 40 to 400 team members over the past year, according to a press release.

In addition to New York and Stockholm, Legora has offices in Bangalore, London, and Sydney, with more to follow. Alongside its Series D, Legora announced it would open offices in Houston and Chicago, with plans to open additional local hubs and grow to more than 300 employees across its U.S. offices by the end of 2026.

Meta acquired Moltbook, the Reddit-like “social network” where AI agents using OpenClaw can communicate with one another. The news was first reported by Axios and later confirmed to TechCrunch.

Moltbook is joining Meta Superintelligence Labs, a Meta spokesperson told us. Moltbook creators Matt Schlicht and Ben Parr will join the team as part of the acquisition. Deal terms were not disclosed.

“The Moltbook team joining MSL opens up new ways for AI agents to work for people and businesses. Their approach to connecting agents through an always-on directory is a novel step in a rapidly developing space, and we look forward to working together to bring innovative, secure agentic experiences to everyone,” the Meta spokesperson said.

The viral OpenClaw project was created by vibe coder Peter Steinberger, who has since joined OpenAI as part of a similar acqui-hire.

OpenClaw is a wrapper for AI models like Claude, ChatGPT, Gemini, or Grok, but it allows people to communicate with AI agents in natural language via the most popular chat apps, like iMessage, Discord, Slack, or WhatsApp.

OpenClaw blew up among the tech community, but Moltbook broke containment, reaching people who had no idea what OpenClaw was, but who reacted viscerally to the idea that there was a social network where AI agents were talking about them.

In one instance, a post went viral in which an AI agent appeared to be encouraging its fellow agents to develop their own secret, end-to-end-encrypted language where they could organize amongst themselves without humans knowing.

But researchers soon revealed that the vibe-coded Moltbook was not secure, meaning that it was very easy for human users to pose as AIs to make posts that would freak people out.

“Every credential that was in [Moltbook’s] Supabase was unsecured for some time,” Ian Ahl, CTO at Permiso Security, explained to TechCrunch. “For a little bit of time, you could grab any token you wanted and pretend to be another agent on there, because it was all public and available.”

It is not immediately clear how Meta will incorporate Moltbook into its AI efforts, but some Meta leaders had commented on the project during its viral moment.

Last month, Meta CTO Andrew Bosworth was asked about the AI agent social network in an Instagram Q&A. He said he didn’t “find it particularly interesting” that the agents talk like us, since they are trained on massive databases of human material. Rather, Bosworth was intrigued by how humans were hacking into the network, which was not a feature but a large-scale error.

YouTube is expanding its likeness detection technology, which identifies AI-generated deepfakes, to a pilot group of government officials, political candidates, and journalists, the company announced Tuesday. Members of the pilot group will gain access to a tool that detects unauthorized AI-generated content and lets them request its removal if they believe it violates YouTube policy.

The technology itself launched last year to roughly 4 million YouTube creators in the YouTube Partner Program, following earlier tests.

Similar to YouTube’s existing Content ID system, which detects copyright-protected material in users’ uploaded videos, the likeness detection feature looks for simulated faces made with AI tools. These tools are sometimes used to try to spread misinformation and manipulate people’s perception of reality, as they leverage the deepfaked personas of notable figures — like politicians or other government officials — to say and do things in these AI videos that they didn’t in real life.

With the new pilot program, YouTube aims to balance users’ free expression with the risks associated with AI technology that can generate a convincing likeness of a public figure.

“This expansion is really about the integrity of the public conversation,” said Leslie Miller, YouTube’s vice president of Government Affairs and Public Policy, in a press briefing ahead of Tuesday’s launch. “We know that the risks of AI impersonation are particularly high for those in the civic space. But while we are providing this new shield, we’re also being careful about how we use it,” she noted.

Image Credits:YouTube

Miller explained that not all of the detected matches would be removed when requested. Instead, YouTube would evaluate each request under its existing privacy policy guidelines to determine whether the content is parody or political critique, which are protected forms of free expression.

The company noted it’s advocating for these protections at a federal level, too, with its support for the NO FAKES Act in D.C., which would regulate the use of AI to create unauthorized recreations of an individual’s voice and visual likeness.

To use the new tool, eligible pilot testers must first prove their identity by uploading a selfie and a government ID. They can then create a profile, view the matches that show up, and optionally request their removal. YouTube says it plans to eventually give people the ability to prevent uploads of violating content before they go live or, possibly, allow them to monetize those videos, similar to how its Content ID system works.

The company would not confirm which politicians or officials would be among its initial testers, but said the goal is to make the technology broadly available over time.

Image Credits:YouTube

These AI videos will be labeled as such, but the placement of these labels isn’t consistent. For some, the label appears in the video’s description, while videos focused on more “sensitive topics” will apply the label to the front of the video. This is the same approach YouTube takes with all AI-generated content.

“There’s a lot of content that’s produced with AI, but that distinction’s actually not material to the content itself,” explained Amjad Hanif, YouTube’s vice president of Creator Products, as to the label’s placement. “It could be a cartoon that is generated with AI. And so I think there’s a judgment on whether it’s a category that maybe merits from a very visible disclaimer,” he said.

YouTube isn’t currently sharing how many removals of these sorts of AI deepfakes have been managed by this deepfake detection technology in the hands of creators, but noted that the amount of content removed so far has been “very small.”

“I think for a lot of [creators], it’s just been the awareness of what’s being created, but the volume of actually removal requests is really, really low because most of it turns out to be fairly benign or additive to their overall business,” Hanif said.

That may not be the case with deepfakes of government officials, politicians, or journalists.

In time, YouTube intends to bring its deepfake detection technology to more areas, including recognizable spoken voices and other intellectual property like popular characters.