Linux, a free, open-source operating system, runs most of the internet’s servers, yet it remains a niche choice on consumer PCs. For years, Linux’s desktop market share in North America generally hovered in the low single digits. However, reports began popping up earlier this week that Linux’s market share had doubled, from the usual 3% to 5% to around 10%, according to Statcounter. And it happened quickly — in just a few weeks between May and June.

Statcounter, which provides usage stats helpful to webmasters and cybersecurity professionals, measures website visits, including the operating system used to access the website. Statcounter can’t tell if it’s a human or a bot visiting a website, but the Linux uptick was corroborated by data from Cloudflare, an internet network company that provides many web-based services, such as content delivery and cybersecurity.

Cloudflare senior product manager Lai Yi Ohlsen told CNET via email that the company’s data, which measures HTTP requests by operating system, shows Linux accounted for 9.8% of US-based internet traffic from July 4 to Aug. 1, closely aligning with the Statcounter metric.

Cloudflare’s data suggests the jump isn’t primarily due to more Americans browsing the web on Linux, however. When the company applies its “likely human” filter to the same July 4 to Aug. 1 dataset, Linux’s share falls from 9.8% to 6.4% — a figure much closer to the operating system’s typical usage levels.

One explanation for such a sharp increase over a short period is the growing use of AI agents. AI agents are bots that automatically do things on your behalf with minimal oversight.

“Linux is the OS of choice for AI agents,” Linux Foundation CEO Jim Zemlin told CNET.

In June, Cloudflare CEO Matthew Prince said bots now generate more web traffic than humans. If that’s the case, shifts in automated traffic could significantly affect operating system usage metrics derived from web requests.

“Welp, that happened faster than I predicted,” Prince said in a post on X on June 3. “Thought it would be [at the] end of 2027, then early 2027, but agentic traffic [is] growing so fast that bots have now passed human traffic online for the first time in the internet’s history.”

The possibility of one or more large web services dramatically increasing their Linux-based activity, such as data scraping, automated crawling or cloud infrastructure workloads, is likely also a factor.

Julian is a contributor and former staff writer at CNET.

See full bio

Most encrypted USB drives announce themselves the moment they connect. The operating system sees scrambled data or an unusual volume that immediately suggests something valuable sits behind a lock. Phantom Drive takes a different approach. Ryan Walker of Rootkit Labs built an open-source flash drive that presents itself as an ordinary 8 GB stick. Disk utilities show no extra space, no encrypted headers, and the remaining capacity is completely hidden unless a specific procedure is taken to unlock it.

Plug the device into any computer and it mounts like a plain, unencrypted drive. You can store everyday files on that 8 GB public area without issue. The firmware watches every write that arrives over USB. When it spots the string “password:” followed by a passphrase in any text file, it extracts the password into memory, wipes the string from the data about to be written so the secret never lands on the card, derives an encryption key, then unmounts the public volume and remounts the remaining space as a new drive. All reads and writes to that second area now pass through AES-256 encryption handled by the hardware itself.

The WCH CH569W, a small RISC-V microcontroller with all the bells and whistles, including a USB interface, SD card controller, and a serious piece of encryption cracking hardware known as an AES acceleration block, is at the core of this whole operation. The storage is provided by a microSD card that you purchase yourself, with the public 8 GB located at the beginning of the card, since the remainder of the card is treated as the secure region once unlocked. The specific salt used in encryption is associated with a unique identification on each board, thus an SD card removed from one Phantom drive will not open in another. Even if you take the card out and insert it into another device, the data will remain scrambled and unusable until you enter the correct password and salt.

Key generation is based on a piece of code called PBKDF2-HMAC-SHA256; you may pick between 100 thousand and 600 thousand iterations, with the lesser setting unlocking you in about 30 seconds and the higher setting delaying it by a few minutes but making it more difficult to guess offline. In terms of encryption mode, you can pick between AES-CTR and AES-XTS, with AES mode providing roughly 20mb/s read and 9mb/s write speeds, while XTS mode is slower. Upgrading the firmware is simple; simply hold down a button while plugging in the drive, and the USB bootloader on the chip will handle the rest.

Everything about this is open source, which means the schematics, board files, firmware source, and even the mechanical designs for the box are all available on GitHub under an MIT license. Walker has already built a limited batch of them, and you can get one from the rootkit labs shop if you don’t mind buying your own microSD card. If you wish to seal the box for extra security, you can use epoxy. However, if you seal it and then try to open it again, you will break the case, but the data on the card will remain safe.
[Source]

Samsung did something genuinely interesting at Galaxy Unpacked 2026. For the first time in the history of its Galaxy Z foldable lineup, it split the book-style foldable line into two entirely different devices. Whether it had something to do with Apple’s purported iPhone Ultra is a conversation for another time, but for now, anyone visiting a Samsung experience center has two Fold phones to choose from: the new Galaxy Z Fold 8 and the Galaxy Z Fold 8 Ultra. 

The Fold 8 is wider, lighter in hand, and built specifically for one-hand use on the cover screen and content consumption on the inner screen. The Fold 8 Ultra, on the other hand, carries forward the tall and thin design. After spending some hands-on time with both devices, I’ve come to a conclusion.

While I genuinely appreciate what the Fold 8 is trying to do, I won’t buy it with my own money. I’ll spend the extra $200 on the Ultra, and I wouldn’t think twice about it. Here’s why.

When a Fold starts feeling like a phone

As I was reading the embargoed spec sheets of both phones, one thing caught me off guard. Even though the Fold 8 weighs 201 grams, which basically makes it the lightest Samsung Fold ever, it’s not the thinnest foldable. That honor belongs to the Ultra. 

The Fold 8 sits at 4.5mm when unfolded, while the Fold 8 Ultra measures 4.1mm. Folded, the Ultra is 8.9mm against the Fold 8’s 9.7mm. You might think that I’m making a fuss over a laughably small difference, but the moment you hold both devices in hand, you’ll understand what I’m saying. 

To me, the Ultra feels flatter, sleeker, and more refined. I strongly believe that Samsung has managed to shave off the bulk that usually comes with a book-style foldable. At 8.9mm, the phone sits more naturally in the palm, feels less chunky, and far less brick-like.

I know we already have conventional flagships breaching the 8mm mark. The S26 Ultra is only 7.9mm thick, which already gives it a remarkably slim yet balanced in-hand feel, but the fact that the Fold 8 Ultra is almost as thick as the iPhone 17 Pro or a mere 0.4mm thicker than the Pixel 10 Pro XL says that it’s already operating in the same physical territory as regular flagship phones.

I’m perfectly aware that some Chinese foldables are thinner, but you can’t buy them in the US, at least not through official channels. They simply aren’t an option for the average American consumer, so I’m putting them out of the question.

The 8-inch inner screen is where One UI 9 shines

I’ve run One UI 9’s split-screen setup on both phones, and I have some very strong thoughts about it. Yes, the Fold 8’s 4:3 inner display is a great screen for watching content. But as a tool for getting some serious work done, it comes up short.

To me, the landscape split-screen option on the 7.6-inch Fold 8 feels slightly cramped vertically. It looks like two browser windows fighting for elbow room on a small desk. The Ultra’s 8-inch display, with its near-1:1 aspect ratio, provides a meaningfully larger vertical canvas to work with.

As someone who has to go through a couple dozen emails every single day, read as many words as fast as possible in news articles on Chrome, and generally squeeze as much information onto the screen as possible, that’s the difference between an annoyingly short and a genuinely usable split-screen setup. 

Every additional line of text directly means less scrolling, fewer page turns, and less time spent moving things around.

Fun fact: the regular Fold 8 lets you use three apps in a vertical split-screen setup side by side, but the individual app windows get way too tight in my opinion. The standard one-plus-two layout works the best for me, and the Fold 8 Ultra’s display plays perfectly into that setup.

If I’m buying a foldable, I’m going to use the additional screen estate to its best, and for that, the Ultra remains the better one. What’s the point of carrying around a tablet-sized inner display if I’m constantly fighting for space on it anyway?

The dedicated telephoto wins me over

I’m not sure about everyone, but I love to capture portraits on a smartphone. And I can’t stress this enough, but the kind of subject magnification a dedicated telephoto camera adds is simply unmatched. No in-sensor 2x crop can match that, not the one on my iPhone 17, and neither the one on the Fold 8. 

To be completely honest with you, I’m not as drawn to the 200MP (f/1.7) camera on the Fold 8 Ultra as I am to the presence of a 10MP (f/2.4) 3x telephoto lens. It’s not the best one out there: the 1/3.94-inch sensor doesn’t do well in poorly lit scenarios, and the night mode just makes handheld photography more tedious. 

However, in broad daylight, or even with some artificial light, the results are much better than anything you can achieve with a digital crop on a 50MP or 200MP primary camera. Concerts and sports events aside, the 3x telephoto, combined with the portrait mode, gives genuinely great results. 

Eventually, I’d want Samsung to improve the 10MP sensor camera with a slightly larger sensor. But for now, spending $1,899 on a 2026 flagship foldable with no zoom camera is a compromise I’m not willing to make. 

Ultimately, that’s what it comes down to for me. The Fold 8 isn’t a bad phone by any measure. Samsung has made the right changes for the right audience: the ones who value content consumption and one-handed usage over everything else. However, if I’m spending nearly $2,000 on a foldable, I want the one that makes the fewest compromises. 

I want the thinner, more premium design, the larger inner screen for productivity-based workflows, and a dedicated telephoto camera. At $200 cheaper, the Fold 8 might be the smarter and more accessible one for some people, but I’d rather spend the difference and get the Ultra for me. 

RayRod TV has released a free VR mod that drops Mario Kart 64 into modern headsets with full head tracking and stereo 3-D. Built on HarbourMasters’ SpaghettiKart PC port, the project runs the original 1997 racing game natively on Windows while adding an OpenXR layer so any compatible headset can take over the view. You supply your own US .z64 ROM, point the executable at it once, and the rest is ready to go.

Head tracking is at the center of this experience. In first-person mode, you’re essentially right in the driver’s seat, with the camera focusing on every movement of your head. Lean in on a curve on Rainbow Road, give the left side a quick glimpse for an upcoming item box, or look up and forward to see the track fly into the distance, since it’s all in your head. The developer took pity on us and provided some ease-back and flip options to prevent the camera from going too far in on drifts. Otherwise, you’d be feeling quite green around the gills after only one lap. Even with those softening settings in place, the impression of speed remains extremely acute. Take a traditional shortcut, such as the tunnel on Koopa Troopa Beach, where the rock ceiling is right on top of your helmet, making it feel like a real life close call.

No products found.

If you need some extra options to get settled in, there are four camera modes to pick from. You’re literally in the kart, so first-person is straight in your face. Third-person mode returns to the usual chase camera, although you can still lean about to see your surroundings. Then there’s Theater mode, which just displays the original flat game on a large virtual screen floating in the empty area. If you want to go all out, Diorama mode reduces the entire course to a little tabletop model, so Rainbow Road simply hangs in the air like a kid’s toy set that you can walk around. Each mode is readily switchable; simply pull out the floating menu when paused and push the appropriate trigger.

That same option allows you a lot of fine control over how the entire experience feels to you. There are separate sliders for world scale and stereo depth, allowing you to adjust the track size and 3-D effect separately if desired. You may change the eye height, camera distance, HUD opacity, and whether the HUD is locked to the world or fixed to your head. If you want to go fully minimalist, you can disable the interface and only see the road before of you. Motion controllers receive all of the rumbling feel on hits, but regular gamepads are also supported, so if you’re a purist, stick with what you know.

Aside from the VR excitement, the update also allows you to customize the racing choices. You have a custom menu where you can stack new rules on top of existing classes. Item behavior can be completely random, or frantic and inverted if you wish. There are also optional high-resolution texture packs that can be dropped in, such as the MK64 Reloaded pack, which sharpens up existing objects without affecting the overall appearance and feel. Everything runs on Windows, but internet multiplayer has yet to be resolved, thus local multiplayer is still the only option for the time being.

Researchers found a way to bypass recent mitigations for Spectre v2 speculative execution side-channel attacks and developed an exploit to leak secrets from Linux machines.

​The method works against Spectre v2 defenses on AMD and Intel processors that rely on sanitizing or isolating branch predictors, which researchers generically refer to as neutralization-based mitigations.

Spectre v2 is also known as Branch Target Injection (BTI) and is a variant of the Spectre class of vulnerabilities.

It exploits a processor’s indirect branch predictor and causes it to mispredict the target of an indirect branch, leading to speculative execution along an attacker-influenced code path.

Modern processors use branch prediction to guess the most likely execution path and speculative execution to run instructions along that predicted path before the branch outcome is known.

Spectre v2 allows an attacker to manipulate the CPU’s indirect branch predictor so that the processor speculatively executes instructions at an attacker-chosen location, which could expose sensitive data.

With neutralization-based mitigations (eIBRS on Intel and Safe RET on AMD), there is a gap between the time the branch predictor is isolated and when it is used by the victim branch.

Active Spectre v2 mitigations assume that an attacker cannot use to their advantage the time between cleaning the branch predictor state and using it.

However, the researchers introduced a primitive that enables re-poisoning the CPU’s state after the cleaning but before it is used.

Daniël Trujillo, a PhD student, and associate professor Mengjia Yan of the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) discovered a technique for exploiting this Time-of-Neutralization to Time-of-Use (TONTOU) window to extract sensitive data.

“An attacker without any special access to read arbitrary memory from the system, including sensitive data such as hashed passwords,” Trujillo told BleepingComputer.

The researchers developed an Interrupt Injection attack, where “unprivileged user programs can schedule timer interrupts to occur during kernel execution.”

“Therefore, we can force the kernel to be redirected to the interrupt handler and use this handler to poison microarchitectural states within the post-neutralization window,” the researchers explain.

The researchers found that interrupts occurring during the post-neutralization window can be used to poison the processor’s indirect branch predictor, enabling attacks against all types of indirect branches.

Mengjia and Trujillo tested the attack starting from the assumption that an attacker can run arbitrary, unprivileged code on a Linux target machine to leak data from the kernel.

On an AMD Zen 2 host with the latest Spectre v2 mitigations, the two researchers successfully ran through all the stages of a TONTOU attack: neutralization, redirection, poisoning, and the use of the poisoned branch predictors.

Successfully exploiting the issue requires overcoming several obstacles, including redirecting kernel control flow, precisely aligning interrupts with the post-neutralization window, and using the interrupt handler to poison the branch predictor entry associated with the target indirect branch.

The researchers address these challenges through installing ‘timers’ to trigger hardware interrupts, frequent injection of interrupts, and via active and passive poisoning methods.

The attack was tested on both Intel and AMD processors. On an AMD Zen 2 system running Linux version 6.14.0-37-generic with 16GB of RAM, the researchers showed it could leak arbitrary kernel memory at a rate of 5.47 bytes/s and 91.97% accuracy, including the contents of /etc/shadow, which stores password hashes.

​Across 10 test runs, the attack successfully located and extracted the file in five cases, with each attempt taking an average of 18 minutes.

While the attack is also possible on Intel machines, they noted that additional software requirements make the task more complex.

The researchers note that interrupt injection enables attacker-controlled poisoning of the Return Stack Buffer (RSB), causing speculative mispredictions of return targets.

Because passive RSB pollution is less reliable, they combined interrupt injection with Inception, a previously disclosed attack that Daniël Trujillo helped develop.

AMD has published an advisory today noting that the interrupt inject issue “appears to be associated” with how implementation on Linux of the Safe RET mitigation against potential information disclosure attacks.

Daniël Trujillo and Mengjia Yan presented their findings today at the Black Hat USA security conference. They will also share the details at the USENIX Security 2026 between October 27 and 29.

Security teams log 54% of successful attacks and alert on just 14%. The rest move through your environment unseen.

The Picus whitepaper shows how breach and attack simulation tests your SIEM and EDR rules so threats stop slipping by detection.

Advertisement

Get the whitepaper

Your ChatGPT usage limits are about to get loosened if you’re a free user. OpenAI said on Thursday that it is giving free users unlimited text chats, and the default model for paying users is getting upgraded.

Usage limits are one of the biggest annoyances of using AI chatbots and other products. If you don’t pony up for the $20 paid subscription — or if your workplace doesn’t pay for an enterprise plan — you probably run into messages on ChatGPT telling you that you’ve reached your limit for the day. Now, these limits for text-only chats will be removed for free users. Limits on image generation, file uploads and voice tools will still apply.

Free and paying users will have a new default model, GPT-5.6 Luna, the lightweight version in the company’s newest model family. Luna has performed better than the current default, GPT-5.5 Instant, with OpenAI reporting that factual errors were 62% less common with Luna. There’s also going to be a new “Think” button available for free users and subscribers on the $8 per month Go plan. This is supposed to help the AI answer harder questions by taking more time to process an answer.

Plus and Pro users ($20 and $200 per month subscribers, respectively) should also see some improvements when they use GPT-5.6 Sol, the heavyweight new model. OpenAI said it tweaked Sol to be better at everyday tasks, like web research, writing, planning and making decisions. It should give honed or focused answers — less unnecessary info and formatting. There will also be a slider tool where you can increase the amount of technical firepower you want in your prompt, like upgrading from GPT-5.6 high to extra high to pro.

Luna will be the new default model starting this week, while the unlimited chats, new Think button and slider should be released next week. Don’t panic if you don’t see them right away; rollouts can sometimes take a few days.

While unlimited chats might be nice if you use AI for school or work, it could be an issue for some vulnerable groups. We’ve seen more than a few cases (and subsequent lawsuits against AI companies) where people who use AI chatbots excessively develop and exacerbate some mental health issues. In one lawsuit, a mother alleges that OpenAI failed to enact necessary safeguards in order to keep her daughter engaged with ChatGPT. It’s a similar concern that is underpinning several lawsuits against social media companies, saying they had deliberately addictive designs that resulted in harm for young and vulnerable users.

OpenAI says it has added guardrails to the models so they “avoid romantic roleplay, age-restricted challenges and presenting itself as a substitute for real-world relationships” for users under 18.

(Disclosure: Ziff Davis, CNET’s parent company, in 2025 filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)

Katelyn Chedraoui

Reporter 2

Katelyn is a reporter with CNET covering artificial intelligence, including chatbots, image and video generators. Her work explores how new AI technology is infiltrating our lives, shaping the content we consume on social media and affecting the people behind the screens. She graduated from the University of North Carolina at Chapel Hill with a degree in media and journalism. You can reach her at kchedraoui@cnet.com.

See full bio

AI and ML

Agent Plugins 1.0 defines a write-once-run-anywhere container for passing tools and skills across different agent platforms

The AI giants have joined forces to ask the world’s developers to all please write their agentic skills and tools into the same format. In return, coders will be able to move these superpowers easily from one platform to another (such as when they get laid off). It’s designed to be write once, run everywhere.

Vercel posted the initial draft of the Agent Plugins spec on Thursday, with input from Amazon, Cursor, Microsoft, and OpenAI, which all pledged fidelity (and engineers) to the project. It was quickly embraced by the Linux Foundation’s Agentic AI Foundation (AAIF), though it will operate as an independent entity.

Agent Plugins 1.0 packages an AI agent’s skills and tools in a single directory structure. It builds on the AAIF-stewarded MCP protocol and the open Agent Skills standard originally created by Anthropic in 2025.

Agent Skills holds the instructions for how the agent should perform a specific task. MCP provides the connectivity to outside databases, APIs, shell scripts, and other external resources. Together, the trio makes AI capabilities modular so that they are, in theory, write-once-run-anywhere. 

It is still up to the agent clients, or platforms, to provide a user interface, establish security boundaries and create marketplaces where the plugins can be shared. Thus far, VS Code, Cursor, GitHub Copilot, ChatGPT, Codex, and Kiro all support the standard.

File-system-as-configuration

Agent Plugins follows the Next.js practice of using the host computer’s directory format as the structure for defining components.

The spec prescribes creating a top-level directory for each plug-in, with a small manifest and schema. The manifest provides the name, version and schema of the plugin in JSON format. Skills are filed in the skills sub-directory, and the MCP server configuration is stored as an mcp.json file.  

Those are the essential elements, though the plugin authors can also use the namespace directory to store additional extensions. This format “keeps the shared format predictable without requiring it to absorb every client-specific experiment happening across agent products,” noted the AAIF in a blog post. 

Bundling the skills with the MCP configuration makes the package portable, allowing it to be copied and used across any supporting agent client.

The narrow way

Agent developers around the globe are adding all manner of experimental hooks, rules and extensions into their skills. Nonetheless, the keepers of the spec deliberately kept the scope “narrow” for this 1.0 release, in AAIF’s words. The best ones may be considered for future versions of the spec, though the team is not making any promises. Until then, developers can add all these extensions to the namespace and hope for the best, portability-wise.  

Going forward, Agent Plugins will be managed under an open governance model, with an independent technical oversight committee calling the shots, considering input from the public. Currently, engineers from Amazon, Cursor, Microsoft, OpenAI, and Vercel are on the core development team. The spec is licensed under the Creative Commons Attribution 4.0 International License (CC-BY-4.0).

There are other efforts underway to tame the wild frontier of agent development. Google folded its own Agent2Agent (A2A) agentic communications protocol effort into AAIF. Google is also building out a specification for finding tools on the web, called Agentic Resource Discovery.

AAIF has a similar “complementary” project to Agent Plugins, Skills Over MCP, which shares skills by bundling them into the MCP server itself. Essentially, this approach is akin to shipping “the manual with the product,” wrote AAIF Vice President Angie Jones. 

Should MCPs be festooned with skills? Or should each set of skills come with an attached MCP? Let the best approach win! ®

A single adapter could be all you need.

The model of GameCube is important

Igor Bonifacic for Engadget

How to connect a DOL-001 GameCube to an HDTV

Igor Bonifacic for Engadget

What about the DOL-101?

Bonus: Nintendo Switch Online

Nintendo

The company has secured a deal to provide Anthropic with computing capacity for six years, according to a Bloomberg report.

Volta Infra, a new AI infrastructure start-up, has emerged from stealth with a valuation of $2.4bn after completing a seed funding round and Series A round led by Azora, Andreessen Horowitz, Altimeter and Nvidia.

The rounds also saw participation from strategic investors including the family office of Michael Dell and Matter Venture Partners.

Volta, founded earlier this year by Ricard Boada and Sofia Gumuzio, aims to help top AI labs and start-ups access key infrastructure. The company develops, finances, builds and operates AI factories by integrating “institutional capital, powered land, data centres, compute, software and operations under a single platform”.

The months-old start-up raised $300m in venture funding, according to Bloomberg, and also announced a deal with asset company Azora to provide $5bn of financing for the “next generation of AI factories” developed by Volta.

“Compute has become a new infrastructure asset class, with AI models and applications as the verticals built on top,” said Boada. “Every technology revolution has run on a physical layer beneath it.

“We founded Volta because compute should be financed, developed, and commercialised with the principles and scale of infrastructure.”

The start-up also announced that it has secured a $10bn strategic partnership with an unnamed AI lab to develop an AI factory in Norway alongside Bitdeer, a Bitcoin miner that operates data centres.

The project – the first within Volta’s broader development pipeline – will comprise a 133MW deployment powered by Nvidia Vera Rubin systems.

As well as the site in Norway, Volta plans to build US sites in Texas and Wyoming, Boada told Bloomberg. The start-up aims to deploy multiple gigawatts of capacity by 2030.

Bloomberg reported that the unnamed AI lab involved in the $10bn Norway site deal is actually AI giant Anthropic, with the publication citing anonymous sources familiar with the matter. As part of the deal, Anthropic will reportedly have access to Volta’s computing capacity at the site for six years.

Anthropic has recently been making a series of deals to boost its access to vital AI infrastructure and compute.

Just a couple of weeks ago, the company announced a partnership with AMD for 2GW of its latest-generation chips in a bid to boost AI capacity and meet growing demands. The Wall Street Journal reported that the deal between the companies is worth “tens of billions of dollars”.

Anthropic has also made similar deals with Broadcom and Amazon.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

In brief: Flock’s AI-powered license plate readers, which have drawn public scrutiny over surveillance and privacy concerns, are traditionally stationary. However, public records indicate that the company is exploring ways to make its cameras more mobile, including through partnerships with delivery and rideshare companies.

A presentation from Flock, obtained by 404 Media, describes abandoned plans to install the company’s automatic license plate readers on 350,000 vehicles operated by Uber, Lyft, and other delivery companies in Georgia. The plan could have significantly expanded the surveillance capabilities of the already controversial AI-powered cameras.

ALPRs from companies such as Flock and Axon are ostensibly designed to record the license plates of passing vehicles and store the data in a searchable cloud database. Authorities have used the cameras, which currently operate in about 6,000 communities across the US, to track the movements of vehicles suspected of being involved in serious crimes.

However, incidents involving misread plates and police officers using ALPRs to stalk people have contributed to a nationwide backlash. Many communities, including Los Angeles, have since deactivated the cameras. ALPRs are usually installed on traffic lights or roadside poles, but a recently uncovered document outlines a proposed partnership between Flock and dashcam provider Nexar.

The presentation highlights the hundreds of thousands of dashcams in Uber and Lyft rideshare vehicles in Georgia, as well as cameras operated by local businesses and homeowners associations, as potential ALPR platforms. While Flock and Nexar ultimately did not ink the deal, it would have allowed Flock to catalog every vehicle that passed those equipped with cameras. Whether the drivers would have been aware of the surveillance remains unclear.

Authorities have utilized mobile camera networks before. Last year, robotaxi company Waymo admitted that it complies with police warrants for footage from its 360-degree cameras, but only after carefully reviewing the requests. Axon and Motorola also provide ALPRs designed for installation on moving vehicles.

The Flock-Nexar presentation was originally obtained through a public records request by Jason Hunyar, a Georgia-based blogger who reports on Flock’s activities in the state. Previous requests have uncovered records indicating that the company erased its supposedly immutable usage records after an officer observed a girl’s gymnastics room and sold data collected by its cameras.

In one post, Hunyar outlines Flock’s plans to install its cameras on robot dogs that could send data back to their Hangzhou-based manufacturer through pre-installed backdoors. Previous reports also revealed that a major Flock investor supports a company capable of intercepting and altering surveillance footage.