Connect with us

Technology

Inside Mark Zuckerberg’s AI hiring spree

Published

on

Inside Mark Zuckerberg’s AI hiring spree

AI researchers have recently been asking themselves a version of the question, “Is that really Zuck?

As first reported by Bloomberg, the Meta CEO has been personally asking top AI talent to join his new “superintelligence” AI lab and reboot Llama. His recruiting process typically goes like this: a cold outreach via email or WhatsApp that cites the recruit’s work history and requests a 15-minute chat. Dozens of researchers have gotten these kinds of messages at Google alone.

For those who do agree to hear his pitch (amazingly, not all of them do), Zuckerberg highlights the latitude they’ll have to make risky bets, the scale of Meta’s products, and the money he’s prepared to invest in the infrastructure to support them. He makes clear that this new team will be empowered and sit with him at Meta’s headquarters, where I’m told the desks have already been rearranged for the incoming team.

Most of the headlines so far have focused on the eye-popping compensation packages Zuckerberg is offering, some of which are well into the eight-figure range. As I’ve covered before, hiring the best AI researcher is like hiring a star basketball player: there are very few of them, and you have to pay up. Case in point: Zuckerberg basically just paid 14 Instagrams to hire away Scale AI CEO Alexandr Wang.

It’s easily the most expensive hire of all time, dwarfing the billions that Google spent to rehire Noam Shazeer and his core team from Character.AI (a deal Zuckerberg passed on). “Opportunities of this magnitude often come at a cost,” Wang wrote in his note to employees this week. “In this instance, that cost is my departure.”

Advertisement

Zuckerberg’s recruiting spree is already starting to rattle his competitors. The day before his offer deadline for some senior OpenAI employees, Sam Altman dropped an essay proclaiming that “before anything else, we are a superintelligence research company.” And after Zuckerberg tried to hire DeepMind CTO Koray Kavukcuoglu, he was given a larger SVP title and now reports directly to Google CEO Sundar Pichai.

I expect Wang to have the title of “chief AI officer” at Meta when the new lab is announced. Jack Rae, a principal researcher from DeepMind who has signed on, will lead pre-training. Meta certainly needs a reset. According to my sources, Llama has fallen so far behind that Meta’s product teams have recently discussed using AI models from other companies (although that is highly unlikely to happen). Meta’s internal coding tool for engineers, however, is already using Claude.

While Meta’s existing AI researchers have good reason to be looking over their shoulders, Zuckerberg’s $14.3 billion investment in Scale is making many longtime employees, or Scaliens, quite wealthy. They were popping champagne in the office this morning.

Then, Wang held his last all-hands meeting to say goodbye and cried. He didn’t mention what he would be doing at Meta. I expect his new team will be unveiled within the next few weeks after Zuckerberg gets a critical number of members to officially sign on.

Tim Cook.
Getty Images / The Verge
Advertisement

Apple is accustomed to being on top of the tech industry, and for good reason: the company has enjoyed a nearly unrivaled run of dominance.

After spending time at Apple HQ this week for WWDC, I’m not sure that its leaders appreciate the meteorite that is heading their way. The hubris they display suggests they don’t understand how AI is fundamentally changing how people use and build software.

Heading into the keynote on Monday, everyone knew not to expect the revamped Siri that had been promised the previous year. Apple, to its credit, acknowledged that it dropped the ball there, and it sounds like a large language model rebuild of Siri is very much underway and coming in 2026.

The AI industry moves much faster than Apple’s release schedule, though. By the time Siri is perhaps good enough to keep pace, it will have to contend with the lock-in that OpenAI and others are building through their memory features. Apple and OpenAI are currently partners, but both companies want to ultimately control the interface for interacting with AI, which puts them on a collision course.

Apple’s decision to let developers use its own, on-device foundational models for free in their apps sounds strategically smart, but unfortunately, the models look far from leading. Apple ran its own benchmarks, which aren’t impressive, and has confirmed a measly context window of 4,096 tokens. It’s also saying that the models will be updated alongside its operating systems — a snail’s pace compared to how quickly AI companies move.

Advertisement

I’d be surprised if any serious developers use these Apple models, although I can see them being helpful to indie devs who are just getting started and don’t want to spend on the leading cloud models. I don’t think most people care about the privacy angle that Apple is claiming as a differentiator; they are already sharing their darkest secrets with ChatGPT and other assistants.

Some of the new Apple Intelligence features I demoed this week were impressive, such as live language translation for calls. Mostly, I came away with the impression that the company is heavily leaning on its ChatGPT partnership as a stopgap until Apple Intelligence and Siri are both where they need to be.

AI probably isn’t a near-term risk to Apple’s business. No one has shipped anything close to the contextually aware Siri that was demoed at last year’s WWDC. People will continue to buy Apple hardware for a long time, even after Sam Altman and Jony Ive announce their first AI device for ChatGPT next year. AR glasses aren’t going mainstream anytime soon either, although we can expect to see more eyewear from Meta, Google, and Snap over the coming year.

In aggregate, these AI-powered devices could begin to siphon away engagement from the iPhone, but I don’t see people fully replacing their smartphones for a long time. The bigger question after this week is whether Apple has what it takes to rise to the occasion and culturally reset itself for the AI era.

I would have loved to hear Tim Cook address this issue directly, but the only interview he did for WWDC was a cover story in Variety about the company’s new F1 movie.

Advertisement
  • AI agents are coming. I recently caught up with Databricks CEO Ali Ghodsi ahead of his company’s annual developer conference this week in San Francisco. Given Databricks’ position, he has a unique, bird’s-eye view of where things are headed for AI. He doesn’t envision a near-term future where AI agents completely automate real-world tasks, but he does predict a wave of startups over the next year that will come close to completing actions in areas such as travel booking. He thinks humans will need (and want) to approve what an agent does before it goes off and completes a task. “We have most of the airplanes flying automated, and we still want pilots in there.”
  • Buyouts are the new normal at Google. That much is clear after this week’s rollout of the “voluntary exit program” in core engineering, the Search organization, and some other divisions. In his internal memo, Search SVP Nick Fox was clear that management thinks buyouts have been successful in other parts of the company that have tried them. In a separate memo I saw, engineering exec Jen Fitzpatrick called the buyouts an “opportunity to create internal mobility and fresh growth opportunities.” Google appears to be attempting a cultural reset, which will be a challenging task for a company of its size. We’ll see if it can pull it off.
  • Evan Spiegel wants help with AR glasses. I doubt that his announcement that consumer glasses are coming next year was solely aimed at AR developers. Telegraphing the plan and announcing that Snap has spent $3 billion on hardware to date feels more aimed at potential partners that want to make a bigger glasses play, such as Google. A strategic investment could help insulate Snap from the pain of the stock market. A full acquisition may not be off the table, either. When he was recently asked if he’d be open to a sale, Spiegel didn’t shut it down like he always has, but instead said he’d “consider anything” that helps the company “create the next computing platform.”

If you haven’t already, don’t forget to subscribe to The Verge, which includes unlimited access to Command Line and all of our reporting.

As always, I welcome your feedback, especially if you’re an AI researcher fielding a juicy job offer. You can respond here or ping me securely on Signal.

Technology

The DJI Romo robovac had security so poor, this man remotely accessed thousands of them

Published

on

The DJI Romo robovac had security so poor, this man remotely accessed thousands of them

Sammy Azdoufal claims he wasn’t trying to hack every robot vacuum in the world. He just wanted to remote control his brand-new DJI Romo vacuum with a PS5 gamepad, he tells The Verge, because it sounded fun.

But when his homegrown remote control app started talking to DJI’s servers, it wasn’t just one vacuum cleaner that replied. Roughly 7,000 of them, all around the world, began treating Azdoufal like their boss.

He could remotely control them, and look and listen through their live camera feeds, he tells me, saying he tested that out with a friend. He could watch them map out each room of a house, generating a complete 2D floor plan. He could use any robot’s IP address to find its rough location.

“I found my device was just one in an ocean of devices,” he says.

A map like the one I saw, with robots and packets trickling in.
Image: Gonzague Dambricourt
Advertisement

On Tuesday, when he showed me his level of access in a live demo, I couldn’t believe my eyes. Ten, hundreds, thousands of robots reporting for duty, each phoning home MQTT data packets every three seconds to say: their serial number, which rooms they’re cleaning, what they’ve seen, how far they’ve traveled, when they’re returning to the charger, and the obstacles they encountered along the way.

I watched each of these robots slowly pop into existence on a map of the world. Nine minutes after we began, Azdoufal’s laptop had already cataloged 6,700 DJI devices across 24 different countries and collected over 100,000 of their messages. If you add the company’s DJI Power portable power stations, which also phone home to these same servers, Azdoufal had access to over 10,000 devices.

Azdoufal says he could remote-control robovacs and view live video over the internet.

Azdoufal says he could remote-control robovacs and view live video over the internet.

When I say I couldn’t believe my eyes at first, I mean that literally. Azdoufal leads AI strategy at a vacation rental home company; when he told me he reverse engineered DJI’s protocols using Claude Code, I had to wonder whether AI was hallucinating these robots. So I asked my colleague Thomas Ricker, who just finished reviewing the DJI Romo, to pass us its serial number.

With nothing more than that 14-digit number, Azdoufal could not only pull up our robot, he could correctly see it was cleaning the living room and had 80 percent battery life remaining. Within minutes, I watched the robot generate and transmit an accurate floor plan of my colleague’s house, with the correct shape and size of each room, just by typing some digits into a laptop located in a different country.

Here are two maps of Thomas’ living space. Above is what we pulled from DJI’s servers without authentication; below is what the owner sees on their own phone.
Screenshots by The Verge

Here’s a fuller floor plan from Gonzague Dambricourt, who tried out a read-only version of Azdoufal’s tool.
Image: Gonzague Dambricourt (X)

Separately, Azdoufal pulled up his own DJI Romo’s live video feed, completely bypassing its security PIN, then walked into his living room and waved to the camera while I watched. He also says he shared a limited read-only version of his app with Gonzague Dambricourt, CTO at an IT consulting firm in France; Dambricourt tells me the app let him remotely watch his own DJI Romo’s camera feed before he even paired it.

Advertisement

Azdoufal was able to enable all of this without hacking into DJI’s servers, he claims. “I didn’t infringe any rules, I didn’t bypass, I didn’t crack, brute force, whatever.” He says he simply extracted his own DJI Romo’s private token — the key that tells DJI’s servers that you should have access to your own data — and those servers gave him the data of thousands of other people as well. He shows me that he can access DJI’s pre-production server, as well as the live servers for the US, China, and the EU.

DJI has MQTT servers associated with the US, EU, and China. I’m not sure what VG stands for.

DJI has MQTT servers associated with the US, EU, and China. I’m not sure what VG stands for.
Screenshot by Sean Hollister / The Verge

Here’s the good news: On Tuesday, Azdoufal was not able to take our DJI Romo on a joyride through my colleague’s house, see through its camera, or listen through its microphone. DJI had already restricted that form of access after both Azdoufal and I told the company about the vulnerabilities.

And by Wednesday morning, Azdoufal’s scanner no longer had access to any robots, not even his own. It appears that DJI has plugged the gaping hole.

But this incident raises serious questions about DJI’s security and data practices. It will no doubt be used to help retroactively justify fears that led to the Chinese dronemaker getting largely forced out of the US. If Azdoufal could find these robots without even looking for them, will it protect them against people with intent to do harm? If Claude Code can spit out an app that lets you see into someone’s house, what keeps a DJI employee from doing so? And should a robot vacuum cleaner have a microphone? “It’s so weird to have a microphone on a freaking vacuum,” says Azdoufal.

It doesn’t help that when Azdoufal and The Verge contacted DJI about the issue, the company claimed it had fixed the vulnerability when it was actually only partially resolved.

Advertisement

“DJI can confirm the issue was resolved last week and remediation was already underway prior to public disclosure,” reads part of the original statement provided by DJI spokesperson Daisy Kong. We received that statement on Tuesday morning at 12:28PM ET — about half an hour before Azdoufal showed me thousands of robots, including our review unit, reporting for duty.

Not just robovacs — DJI’s power stations also use this system.

Not just robovacs — DJI’s power stations also use this system.
Screenshot by Sean Hollister / The Verge

To be clear, it’s not surprising that a robot vacuum cleaner with a smartphone app would phone home to the cloud. For better or for worse, users currently expect those apps to work outside of their own homes. Unless you’ve built a tunnel into your own home network, that means relaying the data through cloud servers first.

But people who put a camera into their home expect that data to be protected, both in transit and once it reaches the server. Security professionals should know that — but as soon as Azdoufal connected to DJI’s MQTT servers, everything was visible in cleartext. If DJI has merely cut off one particular way into those servers, that may not be enough to protect them if hackers find another way in.

Unfortunately, DJI is far from the only smart home company that’s let people down on security. Hackers took over Ecovacs robot vacuums to chase pets and yell racist slurs in 2024. In 2025, South Korean government agencies reported that Dreame’s X50 Ultra had a flaw that could let hackers view its camera feed in real time, and that another Ecovacs and a Narwal robovac could let hackers view and steal photos from the devices. (Korea’s own Samsung and LG vacuums received high marks, and a Roborock did fine.)

It’s not just vacuums, of course. I still won’t buy a Wyze camera, despite its new security ideas, because that company tried to sweep a remote access vulnerability under the rug instead of warning its customers. I would find it hard to trust Anker’s Eufy after it lied to us about its security, too. But Anker came clean, and sunlight is a good disinfectant.

Advertisement

DJI is not being exceptionally transparent about what happened here, but it did answer almost all our questions. In a new statement to The Verge via spokesperson Daisy Kong, the company now admits “a backend permission validation issue” that could have theoretically let hackers see live video from its vacuums, and it admits that it didn’t fully patch that issue until after we confirmed that issues were still present.

Here’s that whole statement:

DJI identified a vulnerability affecting DJI Home through internal review in late January and initiated remediation immediately. The issue was addressed through two updates, with an initial patch deployed on February 8 and a follow-up update completed on February 10. The fix was deployed automatically, and no user action is required.

The vulnerability involved a backend permission validation issue affecting MQTT-based communication between the device and the server. While this issue created a theoretical potential for unauthorized access to live video of ROMO device, our investigation confirms that actual occurrences were extremely rare. Nearly all identified activity was linked to independent security researchers testing their own devices for reporting purposes, with only a handful of potential exceptions.

The first patch addressed this vulnerability but had not been applied universally across all service nodes. The second patch re-enabled and restarted the remaining service nodes. This has now been fully resolved, and there is no evidence of broader impact. This was not a transmission encryption issue. ROMO device-to-server communication was not transmitted in cleartext and has always been encrypted using TLS. Data associated with ROMO devices, such as those in Europe, is stored on U.S.-based AWS cloud infrastructure.

DJI maintains strong standards for data privacy and security and has established processes for identifying and addressing potential vulnerabilities. The company has invested in industry-standard encryption and operates a longstanding bug bounty program. We have reviewed the findings and recommendations shared by the independent security researchers who contacted us through that program as part of our standard post-remediation process. DJI will continue to implement additional security enhancements as part of its ongoing efforts.

Advertisement

Azdoufal says that even now, DJI hasn’t fixed all the vulnerabilities he’s found. One of them is the ability to view your own DJI Romo video stream without needing its security pin. Another one is so bad I won’t describe it until DJI has more time to fix it. DJI did not immediately promise to do so.

And both Azdoufal and security researcher Kevin Finisterre tell me it’s not enough for the Romo to send encrypted data to a US server, if anyone inside that server can easily read it afterward. “A server being based in the US in no way, shape, or form prevents .cn DJI employees from access,” Finisterre tells me. That seems evident, as Azdoufal lives in Barcelona and was able to see devices in entirely different regions.

“Once you’re an authenticated client on the MQTT broker, if there are no proper topic-level access controls (ACLs), you can subscribe to wildcard topics (e.g., #) and see all messages from all devices in plaintext at the application layer,” says Azdoufal. “TLS does nothing to prevent this — it only protects the pipe, not what’s inside the pipe from other authorized participants.”

When I tell Azdoufal that some may judge him for not giving DJI much time to resolve the issues before going public, he notes that he didn’t hack anything, didn’t expose sensitive data, and isn’t a security professional. He says he was simply livetweeting everything that happened while trying to control his robot with a PS5 gamepad.

“Yes, I don’t follow the rules, but people stick to the bug bounty program for money. I fucking don’t care, I just want this fixed,” he says. “Following the rules to the end would probably make this breach happen for a way longer time, I think.”

Advertisement

He doesn’t believe that DJI truly discovered these issues by itself back in January, and he’s annoyed the company only ever responded to him robotically in DMs on X, instead of answering his emails.

But he is happy about one thing: He can indeed control his Romo with a PlayStation or Xbox gamepad.

Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.

Continue Reading

Technology

World’s fastest humanoid robot runs 22 MPH

Published

on

World’s fastest humanoid robot runs 22 MPH

NEWYou can now listen to Fox News articles!

A full-size humanoid robot just ran faster than most people will ever sprint. 

Chinese robotics firm MirrorMe Technology has unveiled Bolt, a humanoid robot that reached a top speed of 22 miles per hour during real-world testing. This was not CGI or a computer simulation. The footage, shared by the company on X, shows a real humanoid robot running at full speed inside a controlled testing facility.

That milestone makes Bolt the fastest running humanoid robot of its size ever demonstrated outside computer simulations. For robotics, this is a line-crossing moment.

Sign up for my FREE CyberGuy Report
Get my best tech tips, urgent security alerts and exclusive deals delivered straight to your inbox. Plus, you’ll get instant access to my Ultimate Scam Survival Guide – free when you join my CYBERGUY.COM newsletter.

Advertisement

WARM-SKINNED AI ROBOT WITH CAMERA EYES IS SERIOUSLY CREEPY

MirrorMe Technology’s humanoid robot Bolt reaches 22 mph during a real-world sprint test inside a controlled facility. (Zhang Xiangyi/China News Service/VCG via Getty Images)

What allows the world’s fastest humanoid robot to run at 22 mph

In the promotional video, the run is shown using a split-screen view. On one side of the screen, Wang Hongtao, the founder of MirrorMe Technology, runs on a treadmill. On the other side, Bolt runs under the same conditions. The comparison makes the difference clear. As the pace increases, Wang struggles to keep up and eventually gives up, while Bolt continues running smoothly, maintaining balance as its stride rate increases.

Bolt takes shorter strides than a human runner but makes up for it with a much faster stride rhythm. That faster rhythm helps the robot stay stable as it accelerates. Engineers say this performance reflects major progress in humanoid locomotion control, dynamic balance and high-performance drive systems. Speed is impressive. Speed with control is the real achievement.

The humanoid robot design choices behind Bolt’s speed

Bolt stands about 5 feet, 7 inches tall and weighs roughly 165 pounds, putting it close to the size and mass of an average adult human. MirrorMe says that similarity is intentional. The company describes this as the ideal humanoid form. 

Advertisement

Rather than oversized limbs or exaggerated mechanics, Bolt relies on newly designed joints paired with a fully optimized power system. The goal is to replicate natural human motion while staying stable at extreme speeds. That combination is what sets Bolt apart.

HUMANOID ROBOTS ARE GETTING SMALLER, SAFER AND CLOSER

MirrorMe says Bolt’s 22 mph run highlights stability and control, not just raw speed. ( Cui Jun/Beijing Youth Daily/VCG via Getty Images)

Why Bolt’s sprint reflects years of robotics development

Bolt did not appear overnight. MirrorMe has focused on robotic speed as a long-term priority since 2016. Last year, its Black Panther II robot stunned viewers by sprinting 328 feet in 13.17 seconds during a live television broadcast in China. Reports suggested the performance exceeded comparable tests involving Boston Dynamics machines. 

In 2025, the company also set a record with a four-legged robot that surpassed 22 mph, reinforcing its focus on acceleration, agility and sustained high-speed motion. China’s interest in robotic athletics continues to grow. Beijing even hosted the first World Humanoid Robot Games, where humanoid robots competed in sprint races on a track.

Advertisement

Why MirrorMe says speed is not the end goal

Running at 22 mph grabs attention, but MirrorMe says speed alone is not the point. The engineers behind Bolt care more about what happens at that speed. Balance, reaction time and control matter more than a headline number. Those skills are what let a humanoid robot move like a trained runner instead of a machine on the verge of tipping over.

That is where the athlete angle comes in. MirrorMe envisions Bolt as a training partner that can run alongside elite athletes, hold a steady pace and push limits without getting tired. By matching and slightly exceeding human performance, the robot could help runners fine-tune form, pacing and endurance while collecting precise motion data. In that context, the sprint is not a stunt. It shows how humanoid robots could move beyond demos and into real training and performance settings.

What this means to you

Humanoid robots that can run at highway speeds are no longer something you only see in demos or concept videos. As these machines get faster and more stable, they start to fit into real-world roles. That includes athletic training, emergency response and physically demanding jobs where speed and endurance make a real difference. At the same time, faster robots bring real concerns. Safety, oversight and clear rules matter even more when machines can move this quickly around people. When robots run this fast, the limits need to be clear.

Take my quiz: How safe is your online security?

Think your devices and data are truly protected? Take this quick quiz to see where your digital habits stand. From passwords to Wi-Fi settings, you’ll get a personalized breakdown of what you’re doing right and what needs improvement. Take my Quiz here: Cyberguy.com.

Advertisement

HUMANOID ROBOT MAKES ARCHITECTURAL HISTORY BY DESIGNING A BUILDING

Engineers say Bolt’s high-speed sprint reflects advances in locomotion control, balance and drive systems. (Photo by Kevin Frayer/Getty Images)

Kurt’s key takeaways

Bolt running at 22 mph is eye-catching, but the speed is not the main takeaway. What matters is what it shows. Robots are starting to move more like people. They can run, adjust and stay upright at speeds that used to knock machines over. That opens the door to real uses, but it also raises real questions. How fast is too fast around people? Who sets the rules? And who is responsible when something goes wrong? The technology is moving quickly. The conversation around it needs to move just as fast.

If humanoid robots can soon outrun and outtrain humans, where should limits be set on how and where they are allowed to operate? Let us know by writing to us at Cyberguy.com.

CLICK HERE TO DOWNLOAD THE FOX NEWS APP 

Advertisement

Sign up for my FREE CyberGuy Report
Get my best tech tips, urgent security alerts and exclusive deals delivered straight to your inbox. Plus, you’ll get instant access to my Ultimate Scam Survival Guide – free when you join my CYBERGUY.COM newsletter.

Copyright 2026 CyberGuy.com. All rights reserved.

Advertisement
Continue Reading

Technology

4chan’s creator says ‘Epstein had nothing to do’ with creating infamous far-right board /pol/

Published

on

4chan’s creator says ‘Epstein had nothing to do’ with creating infamous far-right board /pol/

Epstein had nothing to do with the reintroduction of a politics board to 4chan, nor anything else related to the site. The decision to add the board was made weeks beforehand, and the board was added almost 24 hours prior to a first, chance encounter at a social event. His assistant reached out to me afterward, and I met with him one time for an unmemorable lunch meeting. This happened at a time when I was meeting hundreds of people a month while speaking and networking at tech events.

I did not meet him again nor maintain contact. I regret having ever encountered him at all, and have deep sympathy for all of his victims.

Continue Reading

Trending