A human pancreas typically contains about a billion beta cells, the primary producers of insulin. Among people with diabetes, these cells are either missing or dysfunctional, and thus scientists have been looking for ways to replenish their numbers as a possible long-term treatment or even cure for the condition. A team of researchers recently took a step closer to that reality, genetically altering another type of pancreatic cell to produce and release insulin in response to high blood sugar.

The researchers engineered ductal cells, which—as some previous studies had found—occasionally transform into beta cells all on their own. This is an unusual phenomenon among cells in adult bodies, which tend to be firmly fixed in their identities, and it hinted to the scientists that these ductal cells would be a good place to start.

Scientists previously had no clue what genes were driving this metamorphosis, said Jian Li, a postdoctoral researcher at Harvard Medical School who led the recent study in Science Translational Medicine. The researchers therefore took an approach called a genetic screen, which involves breaking little bits of DNA all across the genome to see which are important for a certain biological process. It’s a bit like individually removing pieces of a car engine without a schematic and seeing what breaks to learn which components are involved in delivering fuel or are essential for steering.

Through this process, the researchers found that taking out a gene called ALDH3B2 turned ductal cells into beta-like cells at a higher rate. Without the genetic alteration, fewer than 1 percent of ductal cells spontaneously took on a beta-cell-like state, but when ALDH3B2 was silenced, that proportion increased to about 8.5 percent.

Those initial experiments were performed on human cells in a dish, which the researchers then transplanted into mice with diabetes. Human insulin began circulating in the mice, and the animals’ glucose levels dropped to near-normal levels. These effects persisted for six weeks.

Science has previously explored some gene therapies for diabetes with the hopes of offering long-term relief. For example, a clinical trial launched earlier this year is taking the creative approach of equipping muscle cells with the genetic instructions to make insulin. Other emerging treatments aim to generate new insulin-producing cells in the laboratory and then transplant them into the patient.

This comes with a few risks, namely activation of the immune system. This new study points to another possibility: harnessing the cells that already exist in the pancreas and causing them to change function by turning off some of the genetic switches that maintain their identity.

That method also has its own hurdles, one of the most important of which is ensuring that only the target cells are edited. ALDH3B2 is used by many cells in the body, not just in the pancreas, so precision is important to prevent unforeseen complications.

There’s also a major lingering question: How is this gene involved in turning ductal cells into beta cells? “That’s the part we need to verify first,” Li said. Then, “the next step is either gene therapy or to find specific small molecules to inhibit this gene to see if we can achieve a similar—or even better—effect.”

Even partial improvement could make a massive impact in the lives of the estimated 830 million people who have diabetes worldwide, including many who die each year due to related complications.

This story originally appeared on WIRED en Español and has been translated from Spanish.

Update: Article rewritten with official confirmation from Citrix.

Citrix has confirmed that two critical NetScaler remote code execution vulnerabilities, tracked as CVE-2026-88771 and CVE-2026-88772, are being exploited in attacks and that it has released security updates to fix the flaws.

The vulnerabilities are the same zero-days that cybersecurity researchers, IT providers, and national cybersecurity agencies began privately warning organizations about over the weekend.

NetScaler appliances are particularly valuable targets because organizations commonly deploy them as Internet-facing edge devices that provide remote access and application delivery services for internal corporate networks.

Compromising one of these devices can give attackers an initial foothold at the perimeter of a victim’s network and potentially provide a path to internal systems without first compromising an endpoint inside the organization.

The first signs of the incident appeared when Citrix administrators began reporting on Reddit that IT suppliers and security teams were privately contacting their organizations and advising them to shut down their NetScaler appliances.

“We got a call from our IT supplier’s security team, they couldn’t give any details but they advised to shut our Netscalers down immediately,” one administrator wrote.

Other administrators said law enforcement, CERTs, and national cybersecurity agencies had also been contacting organizations about the issue.

Cybersecurity firm watchTowr later publicly warned that it was “rapidly reacting to rumors” that multiple unpatched Citrix NetScaler remote code execution vulnerabilities were being exploited in the wild after verifying the information with “authoratitive sources.”

“We are currently rapidly reacting to rumors that multiple unpatched Citrix NetScaler RCE vulnerabilities are circulating in the wild. While details are scarce, the information is credible,” watchTowr said.

Citrix confirms active exploitation

Citrix has now published security bulletin CTX697096, confirming the vulnerabilities and releasing patches for affected NetScaler ADC and NetScaler Gateway appliances.

CVE-2026-88771 is a remote code execution vulnerability caused by improper input validation, allowing an unauthenticated attacker to execute arbitrary commands. It has a severity score of 9.5.

Citrix says the flaw affects all NetScaler ADC and NetScaler Gateway deployments, including those using the default configuration, and does not require any additional feature to be enabled.

CVE-2026-88772 is a memory overflow vulnerability that can lead to remote code execution or a denial-of-service condition, also with a severity score of 9.5.

This vulnerability can be exploited when DTLS is enabled on a NetScaler ADC or NetScaler Gateway. Citrix notes that DTLS is enabled by default on VPN virtual servers.

Citrix has confirmed that both flaws have been exploited in attacks against NetScaler devices as zero-days.

“Exploits of CVE-2026-88771 and CVE-2026-88772 on unmitigated NetScaler deployments have been observed,” Citrix said in the security bulletin.

Citrix says the following versions are affected:

• NetScaler ADC and NetScaler Gateway 14.1 before 14.1-73.37
• NetScaler ADC and NetScaler Gateway 13.1 before 13.1-64.23
• NetScaler ADC FIPS before 14.1-73.37 FIPS
• NetScaler ADC FIPS and NDcPP before 13.1-37.279

Secure Private Access Hybrid deployments using NetScaler instances are also affected and must be upgraded to the recommended builds.

Citrix says the bulletin only applies to customer-managed NetScaler ADC and NetScaler Gateway appliances. Cloud Software Group is upgrading Citrix-managed cloud services and Citrix-managed Adaptive Authentication.

The security bulletin also fixes six other NetScaler vulnerabilities, bringing the total to eight flaws fixed in this update.

NCSC warned organizations before disclosure

Before Citrix publicly disclosed the vulnerabilities, the Dutch National Cyber Security Center (NCSC-NL) reportedly sent a pre-notification to organizations in the Netherlands warning about two critical NetScaler zero-days.

Multiple people shared copies of the notification online, which said the agency had received information from a European partner CERT regarding two vulnerabilities that could independently lead to remote code execution.

According to the notice, one vulnerability allowed attackers to place shellcode directly into memory, while technical details about the second vulnerability were still being researched.

At the time, no CVE identifiers had been assigned, and Citrix had not yet published an advisory.

The notification said Citrix discovered the vulnerabilities while investigating incidents in customer environments and identified active exploitation.

It also said Citrix submitted a notification under the European Union’s Cyber Resilience Act after discovering the attacks.

The NCSC said exploitation had been identified at multiple Citrix customers worldwide, although it did not know whether the attacks were widespread.

The agency also warned that exploitation attempts could increase once Citrix released patches and additional technical details.

Because NetScaler upgrades can cause downtime, the NCSC said the warning was intended to give organizations time to prepare, implement safeguards where possible, and install patches quickly once they became available.

BleepingComputer contacted the Dutch NCSC to confirm whether the advisory circulating online was legitimate.

The agency declined to confirm the notification, saying it could not provide further information to organizations outside its constituency.

“As part of our role as the National CSIRT and sectoral CSIRT for designated organizations, the NCSC-NL monitors relevant developments and cyber threats affecting the Netherlands 24/7,” the NCSC-NL told BleepingComputer.

“We provide information and advice to organizations so that they can take appropriate measures. As you’re not part of our constituency, we cannot disclose any further information at this time.”

Now that Citrix has released fixes and confirmed exploitation, administrators should upgrade affected NetScaler ADC and NetScaler Gateway appliances to the patched versions as soon as possible.

Organizations that cannot apply the updates immediately should reduce Internet exposure where operationally possible until they can patch the appliances.

Join Mikko Hyppönen and security leaders from the NFL, CHANEL, and Atlassian for a two-hour digital summit on what AI-speed attacks change, what defenders should stop doing, and how to validate, decide, fix, and re-validate at machine speed.

Save your seat

Scott Manley teaches orbital mechanics through Kerbal Space Program, and he also DJs. For his latest project, he wired those two lives together on a whim, writing a small Python program that reads MIDI events from a portable mixer and turns them into spacecraft commands. The question he set out to answer was simple enough to print in all caps: can you fly a spaceship using a DJ mixer? After one messy airplane test and one surprisingly clean lunar landing, the answer was yes.

A Numark DJ2Go2 is at the core of this operation, but in a very small way, as this 2-channel USB controller is small enough to carry in a backpack. Oh, and it runs totally on MIDI, so no power issues there. Manley needed to complete a task, so he began mapping inputs that were never intended for pitch or yaw to do so, among other things. He did it all through a config file, which allowed him to adjust the dead zones and response curves without having to start over. As a result, Kerbal’s kRPC remote control interface could accept the translated commands and send them to the 2015 simulator. Throttle originated from a slider on the speed dial that, if yanked the wrong direction, does the opposite of what you’d expect. The attitude was taken care of by the little self-centering sticks and the jog wheels, which always seem like you’re playing with a vinyl record when you only need a few degrees of pitch, such as when you’re descending. Finally, buttons were utilized to control staging, SAS modes, RCS, landing gear, and action groups.

Estes 1469 Tandem X Rocket-Building Kit, Beginner Flying-Rocket Model Kit for Ages 10+, Includes Launch…

• BEGINNER MODEL-ROCKET LAUNCH SET: The Tandem-X rocket-model launch set offers adults and kids ages 10+ hours of fun during the holidays as they…
• 2 SOARING ALTITUDE HEIGHTS: Our Tandem X set offers a high-performing power duo with our giant 30-inch Amazon model (600-foot projected altitude with…
• READY TO ASSEMBLE: Our beginner model-rocket launch set comes with 2 build options. The precolored Amazon model features plastic fins and self-stick…

First up was a plane, and it was a close call because the plane almost won, but Manley quickly discovered that the rotation on the mixer was far too delicate when close to the ground. He nearly crashed the plane into the trees by over-correcting, then spent far too much time battling the camera before eventually figuring it out. At that point, he stopped treating the board as a flight stick and began treating it as a collection of poorly labeled sliders. He described the learning curve as steep and the configuration as unsuitable for low-altitude flight. SAS modes that propel a rocket into space are ineffective when flying a winged plane that follows the terrain. That particular flight landed on the runway due to sheer chance and early saves, rather than design.

When it came to rockets, it was a very different situation. Manley fired one, kept a steady path with the jog wheels, and then let the SAS handle the ride through the first stage burn. It wasn’t until he was climbing that things became a little unstable, and he had to cut the thrust to save the thing from frying itself, before staging, restarting, and sending it on its way. After a few minutes, they were in a 100-kilometer circular orbit, with no drama. From there, he focused on the Mun, decreased the orbit’s lowest point, and began a drop toward a good, smooth crater that would be far superior than the surrounding ridges for landings. Then it was simply a matter of directing it down to the ground with the jog wheels until the lander tipped in and landed on the soil.

Time acceleration was a continual battle for him because a poorly mapped reset caused the clock to spool back and added some annoying visual artifacts, so he saved frequently and flew the coast phases in real time more than he wanted to. Don’t even get started on the radar altitude, which never seemed to appear on the board, thus the last hover was a bit of a gamble, relying on camera zooms and enough practice to overcome the homing inaccuracy. The familiar layout of staging buttons and SAS modes, however, came readily to him, as it was similar to thinking about mixing tracks as a DJ, with distinct cues and a continuous fader. The problem was that planes required a continual stream of micro corrections right near the ground, and rockets provided him with extended burn periods and some nice staging events, but he needed a computer that could at least retain its position while he mixed it all together.
[Source]

A council in north Devon backed a huge AI datacentre plan more than a year before the public knew about it. The documents were reported by Sandra Laville for the Guardian.

Campaigners got the papers through freedom of information requests. They show the developer, Xlinks, had been in private talks with Torridge district council since March 2025. The council sent a letter of support in April 2025, before any environmental study or public consultation.

The council told the Guardian that such early talks are standard and confidential, and do not mean a proposal has been approved.

Xlinks plans a 1.5GW AI data campus and a 1.8GW battery storage site on about 344 hectares near Great Torrington, in a Unesco biosphere reserve. The company puts the cost at up to £13.8bn and says the site will create 600 to 1,200 permanent jobs.

Local campaigners point to the protected countryside, the loss of farmland, water use, noise and fire risks from the batteries. North Devon is also still under a hosepipe ban after a record hot summer.

One document showed that in severe heat and drought, the site could use about 2.9bn litres of water a year. However, Xlinks told the Guardian that figure came from an old design, and that the campus will mainly reuse water and top up from rain.

“How can the largest and most costly infrastructure proposal in the nation be conducted almost exclusively in secret?” said Emily Clark of the Devon Datacentre and BESS action group.

Meanwhile, a national petition started by Devon county councillor Cheryl Cottle-Hunkin has passed 36,000 signatures. It calls for a pause on datacentres and battery sites outside the government’s AI growth zones.

The government rejected the idea on 17 September. It said such projects already face planning and environmental checks, and that battery storage is needed to clean up the power grid by 2030.

Xlinks said it will design the campus with care for local people and nature. A public consultation planned for July has been put back to later this year, the council said.

The fight in Devon follows a wave of datacentre protests in the US, where local groups blocked $130bn of projects in the first three months of this year.

What we know so far: World’s Edge, the Microsoft studio that oversees the Age of Empires franchise, has lost nearly half its staff in the latest round of Xbox layoffs. The cuts coincided with Microsoft’s decision to place the studio under Activision. Former employees also said the studio’s next project was canceled during the transition. The moves raise questions about World’s Edge’s future role and the direction of Age of Empires under Activision.

“Our studio was reorganized to be under the Activision umbrella,” former senior software engineer Andrew Martz said in a LinkedIn post. “As part of that reorg, our next project was canceled and a large chunk of our studio was laid off. I was part of that layoff.”

The layoffs are part of 268 job cuts across Microsoft’s Xbox business. They contribute to the company’s projected target of 3,200 job losses during the current financial year.

World’s Edge has played a different role from a conventional game-development studio. It has overseen the Age of Empires brand and worked with outside developers on new releases, expansions and ports. That setup has allowed Microsoft to maintain several games in the series at once, including remasters of older titles and newer releases for consoles.

– Pirat_Nation 🔴 (@Pirat_Nation) September 22, 2026

Relic Entertainment, for example, was the main developer of Age of Empires IV. World’s Edge helped guide the project as the franchise’s steward. Age of Empires IV arrived on PC in 2021 before later coming to Xbox and PlayStation.

The layoffs mean it’s unclear what happens next for the franchise. Microsoft has not detailed World’s Edge’s plans under Activision or said what the canceled project was. It is also unclear how much of the studio’s work will remain focused on Age of Empires.

Adam Isgreen, World’s Edge’s former studio creative director, said many longtime employees were affected. Isgreen had spent 16 years working on Age of Empires, including his time at World’s Edge.

“I am of course disappointed that this is the way the time with World’s Edge and Age ends for myself and many of my talented teammates that were also impacted – we had some great plans for what’s next,” Isgreen wrote. “I hope the team that remains can rise to the occasion down the line.”

The Age of Empires series has continued to expand beyond its original PC audience. Modern releases have had to support online multiplayer, regular updates, console controls and different hardware platforms, while maintaining the core systems that define the games: resource gathering, base building, large-scale unit control and civilization-specific strategies.

That work is spread across several teams and partners. World’s Edge has been responsible for keeping the franchise’s games and content aligned, even when another studio handled most of the development. A smaller staff could make that job more difficult, especially if Microsoft plans to pursue another major entry in the series.

The reduction comes as new Age of Empires content continues to arrive. Viking Sagas, the latest downloadable expansion for Age of Empires II: Definitive Edition, just launched. It adds three civilizations and 15 single-player campaign scenarios.

Another former World’s Edge employee, lead principal software engineer Greg Snook, said he was informed of his departure while on vacation.

“‘It was nice working with you’ was not a message I expected to get while on vacation,” Snook wrote. “Like many others, I was cut loose from Microsoft today. Turns out my last day with building access is also my exact 25-year hire anniversary.”

For now, the fate of World’s Edge under Activision remains uncertain. The studio still has a major franchise attached to it, but the layoffs and project cancellation suggest its role inside Microsoft’s gaming organization has changed significantly.

The EU’s special envoy for industrial AI has told the EU’s top officials to focus on using AI to lift Europe’s slow productivity. Racing to build the best models matters less, he told them. Zoya Sheftalovich reported the story for Politico.

Jim Hagemann Snabe gave the message to the 27 commissioners at a private seminar on 4 September, two Commission officials told Politico. He said AI could help health, farming, transport, defence and factories.

A week earlier, a Commission AI expert gave a similar briefing to the commissioners’ heads of cabinet. The US and China may lead on models and computing power, but Europe can still gain by putting AI to use, according to slides seen by Politico.

The slides gave examples such as easier testing rules for self-driving cars, secure data spaces for personalised medicine and a digital adviser for farmers.

Commission President Ursula von der Leyen made a similar point in her State of the Union speech on 16 September. She said Europe does not need to build frontier AI to get the most value from it.

Snabe, who chairs the supervisory board of Siemens, was appointed in June to advise von der Leyen and tech chief Henna Virkkunen. His appointment drew conflict-of-interest criticism at the time.

A Commission spokesperson, Thomas Regnier, told Politico the talk was about how Europe can get firms to use AI faster.

However, not everyone backs this plan. German MEP Axel Voss said the Commission was too slow to enforce its AI law, and pointed to recent hacks by AI agents, Politico reported.

Others want Europe to build more of its own AI capacity. Earlier this month, former commissioner Margrethe Vestager and economist Philippe Aghion called for Europe to hold 15% of global compute by 2030.

Snabe’s team will study how each sector can use AI until December, then draft new policies near the end of the year, according to Politico. A final report and strategy are expected in early 2027.

Amazon’s next big sale is still a little over a week away, with Prime Big Deal Days officially running on October 6 and 7. Even then, there are still plenty of Bluetooth speaker deals available for the early buyers, and a few of them are already considerably better than I expected.

I went through the early discounts and narrowed them down to five speakers I’d pay attention to right now. There’s a tiny JBL for less than $40, a Flip 7 sitting well below $100, and options from Bose, Ultimate Ears, and Marshall if you have a little more room in the budget.

Note: Amazon prices and discounts can change quickly and may differ between colors.

JBL Go 5 — $39.95 (27% off)

The JBL Go 5 is the easiest recommendation here if all you really want is a small speaker that can disappear into a backpack. Amazon has it at $39.95, which takes $15 off its $54.95 price. It is tiny at just 0.51 pounds, yet JBL still gives it an IP68 rating for dust and water resistance. Battery life reaches eight hours, with Playtime Boost squeezing another two hours out of it, and the Go 5 also supports Auracast if you eventually want to connect compatible speakers together.

JBL even added ambient edge lighting this generation, which is one of the upgrades over its predecessors At under $40, this is the one I’d throw into a bag for a trip, keep around the bathroom, or take somewhere I wouldn’t necessarily want to risk a more expensive speaker.

Ultimate Ears Wonderboom 4 — $51.44 (49% off)

Spending another $11 gets you something bigger. The blue Ultimate Ears Wonderboom 4 has dropped to $51.44 on Amazon, which is one of the better discounts in this entire list. Amazon commonly sells it below its $99.99 MSRP, yet $51.44 remains an unusually low price and sits only a couple of dollars above its previous Prime Day low.

I also like how it’s built for outdoor use with an IP67 rating. It can even float if it falls into water and survive drops from up to five feet. You get up to 14 hours of audio playback on a single full charge. Its 360-degree design makes placement fairly painless, while Outdoor Boost changes the tuning for open spaces and Podcast Mode focuses more heavily on voices. For just over $50, there’s a lot going on here.

Marshall Emberton III — $144 (20% off)

Marshall goes in a different direction from the utilitarian JBL and Bose designs, and the Emberton III is easily the speaker I’d pick if I cared about how the thing looked sitting on a desk or shelf. The Black/Silver Amazon listing has recently hovered around $144, putting it roughly $36 below Marshall’s $179.99 price.

The bigger reason I’d spend the extra money is battery life. Marshall claims more than 32 hours from a charge, comfortably the longest figure among these five speakers, and a 20-minute quick charge can add another six hours. The Emberton III also has an IP67 rating, Bluetooth 5.3 LE, a built-in microphone for calls, and Marshall’s True Stereophonic multidirectional sound. Two 2-inch 10W full-range drivers sit inside, accompanied by passive radiators.

It is the most expensive pick here. Still, the combination of battery life, ruggedness, and that iconic Marshall design gives it a clear place in this list.

JBL Flip 7 — $119.95 (20% off)

For those looking for a bigger JBL portable speakers, the Flip 7 is the perfect pick. It’s compact enough to carry around without becoming luggage, while giving you considerably more speaker than the tiny Go 5. Amazon currently has the black model at $119.95. JBL rates the Flip 7 for up to 16 hours of playback with Playtime Boost, alongside an IP68 rating that covers water and dust resistance. It is also drop-proof from up to one meter. The newer PushLock system lets you attach interchangeable carrying accessories, and USB-C supports lossless audio when connected to a compatible source.

This makes the Flip 7 the safest all-rounder of these five for me. It has enough battery life and durability for travel, enough output to make sense outdoors, and still occupies very little room in a backpack. At around $120, it sits nicely between the cheaper compact speakers and the more style-driven Marshall.

Beats Pill — $99.95 (33% off)

The Beats Pill has been around in one form or another for years, and the 2024 revival is currently available for under $100. It is currently listed on Amazon as a limited-time deal, with the price sitting at $99.95. Beats claims up to 24 hours from a charge, placing the Pill well ahead of both JBL models above. The USB-C port is useful in both directions too. You can use it for high-resolution lossless audio with a compatible wired source or pull power from the speaker to charge your phone.

Beats also gave the Pill an IP67 rating, a removable lanyard, Apple and Android one-touch pairing, Find My support, and the ability to pair two speakers together in Amplify or Stereo modes. There’s even a microphone for calls and voice assistants. Beats also gave the Pill an IP67 rating, a removable lanyard, Apple and Android one-touch pairing, Find My support, and the ability to pair two speakers together in Amplify or Stereo modes. There’s even a microphone for calls and voice assistants.

A VPN and Tor can both hide your normal public IP address from websites, but they use fundamentally different trust and routing models. A conventional VPN sends covered traffic through one provider-controlled endpoint, while Tor sends traffic through multiple relays so no single ordinary relay normally knows both where the connection originated and which destination it ultimately reaches.

Here, Tor means the Tor network as most people use it through Tor Browser. It does not mean the separate Tor VPN Beta, which the Tor Project currently describes as testing software that should not be relied on for sensitive activity.

VPN vs Tor at a Glance

The easiest way to compare VPN and Tor is to separate routing, trust, browser privacy, and application coverage. Neither option is universally better because they solve overlapping but different privacy problems.

The table shows why “which one hides my IP?” is too narrow a comparison. Both can change the address a destination sees, but the more consequential difference is which parties can observe the user, the route, and the destination.

The Biggest Difference Is Where You Place Trust

A conventional VPN primarily changes the network intermediary you rely on. Your device establishes an encrypted tunnel to a VPN server, and covered traffic exits through that provider’s infrastructure before continuing toward its destination.

Cloudflare’s VPN explanation describes the client-to-server encrypted tunnel and the VPN server’s role as the network endpoint visible to destinations. The useful privacy question is therefore not whether a VPN removes trust, but whether you are comfortable placing that routing role with the VPN operator.

Tor takes a different approach. The Tor Project explains that ordinary Tor traffic passes through multiple relays, with layered encryption and different information available at each position. Its comparison with one-hop proxy services says Tor normally routes traffic through at least three servers before the destination.

For an ordinary web circuit, the first relay is the guard or entry relay. It receives the user’s connection but does not know the final destination. A middle relay connects adjacent parts of the circuit. The exit relay connects to the ordinary public internet but does not receive the user’s original public IP as the source of that exit-side connection.

The Tor client also negotiates separate encryption keys for each hop. The Tor Project’s technical overview of Tor circuits explains that each relay knows only its adjacent hops rather than the complete route.

If you are comparing single-intermediary routing models, the distinction between a proxy and a VPN provides useful context. Tor differs from both by deliberately distributing the route across multiple relays.

What Your ISP, VPN Provider, Tor Relays and Websites Can See

Privacy claims make more sense when you identify the observer instead of asking whether traffic is simply “hidden.”

With a VPN

Your internet service provider still carries the connection to the VPN infrastructure. For traffic actually routed through the VPN, the ISP sees that VPN connection rather than directly carrying each covered connection to its final website.

The VPN operator occupies the next major trust position because traffic reaches its infrastructure before exiting toward internet destinations. A website normally sees the VPN server’s public IP address for traffic leaving through that endpoint.

HTTPS remains a separate protection layer. A VPN tunnel can end at the VPN server while an HTTPS connection remains encrypted between your browser or application and the website.

With Tor

The Tor Project says that when you connect directly to Tor, a local internet provider or other local observer can normally tell that you are communicating with Tor nodes but does not receive the final browsing destination from the normal Tor route. Websites instead see a connection coming from the Tor network rather than from your normal public IP. Tor’s protection overview explains this separation through its multi-hop routing model.

The guard relay sees the incoming client IP because it receives the connection. The exit side knows which ordinary internet destination it connects to. The design goal is that one ordinary relay does not receive both pieces of information as part of normal circuit operation.

This is why statements such as “Tor relays cannot see anything” are inaccurate. Each relay has information required for its role; the protection comes from separating those roles.

A useful follow-up distinction is what your ISP can see with Tor or a VPN, because local-network visibility and destination-side visibility are different parts of the same connection.

Privacy Is Not the Same as Anonymity

Neither VPN routing nor Tor routing prevents you from identifying yourself voluntarily.

If you sign in to an email account connected to your real identity, the service can still associate the session with that account. The same applies if you submit your name, phone number, address, payment details, or other identifying information.

The Tor Project makes this boundary explicit. Its Tor Browser safety guidance says perfect anonymity cannot be guaranteed and explains that signing into a website or supplying personal information identifies you to that service even though Tor may still hide your normal network location.

A conventional VPN has another limitation: its network tunnel does not automatically change the browser environment. Cookies, signed-in sessions, extensions, fonts, screen characteristics, and other browser signals can remain available to websites even when your public IP changes.

IP privacy is therefore only one part of online identifiability.

Tor Browser Adds Protections a VPN Tunnel Does Not

Browser fingerprinting combines observable browser and device characteristics that can help distinguish one browser from others. Signals may include operating-system information, fonts, browser capabilities, extensions, screen characteristics, cookies, and other state.

A VPN tunnel does not standardize those properties. Its primary function is network tunneling and routing.

Tor Browser is designed around a broader browser-privacy model. Tor Project documentation says the browser is modified to prevent or reduce several forms of identifying leakage and to make Tor users look more alike.

The distinction is important enough that the Tor Project strongly recommends against using ordinary browsers as substitutes for Tor Browser. Its current guidance specifically warns about real-IP exposure through DNS or WebRTC, fingerprinting through operating-system details, fonts and plugins, and persistent tracking state such as cookies and cache.

This does not mean Tor Browser makes every user indistinguishable under every circumstance. It means its privacy model includes browser-level defenses that a VPN tunnel by itself does not provide.

The same distinction explains why Tor Browser and private browsing should not be treated as equivalent simply because both can limit some browsing state.

Where Encryption Starts and Stops

Both VPN and Tor comparisons become misleading when “encrypted” is treated as one uninterrupted property from the device to every destination.

VPN encryption

A conventional VPN protects covered traffic inside its tunnel between the VPN client and VPN endpoint. After that point, the application’s own protocol determines what protection continues toward the destination. For ordinary modern web browsing, HTTPS can independently protect the browser-to-website session even though the VPN tunnel itself ends at the VPN server.

Tor and ordinary websites

Tor applies layers of encryption inside its relay circuit, but an ordinary website reached through an exit relay still relies on destination-side encryption such as HTTPS to protect application data on the public-internet side of the route.

The Tor Project’s Tor and HTTPS explanation treats the two protections separately. Tor changes network routing and conceals the user’s normal public IP from the destination, while HTTPS protects data exchanged with an HTTPS website.

Onion services use a different path

An onion service does not use the same public-internet exit pattern. The Tor Project’s onion-service documentation explains that both the client and service establish Tor circuits and meet through a rendezvous point, while onion-service traffic is encrypted between the client and onion host.

A `.onion` service is therefore not simply an ordinary website viewed through a Tor exit. The service connection remains inside Tor rather than emerging through a public exit relay to reach an ordinary public web server.

Tor Distributes Trust, but It Still Has Limits

Tor’s distributed design prevents one ordinary relay from simply receiving the complete source-and-destination relationship, but that is not the same as protection against every possible observer.

The Tor Project explicitly documents traffic-correlation and timing attacks as a remaining limitation. An observer capable of seeing both the user’s side of a connection and the relevant destination or exit side may compare timing patterns. Tor’s documentation on remaining attacks states that Tor does not defend against an observer that can see both ends of the communication channel.

That limitation does not make the relay architecture meaningless. It defines the threat model. Tor still prevents ordinary single relays, destination websites, and local network observers from automatically receiving the complete source-to-destination relationship.

Tor can also lose anonymity through application behavior rather than a break in onion routing. The Tor Project notes that circuits may carry multiple TCP connections, so mixing anonymous and identifying traffic or using applications carelessly can create additional correlation risks.

The more detailed architecture behind Tor guards, middle relays, exit nodes, and onion routing explains why these protections and limitations follow from the network design.

VPN vs Tor for Speed and Everyday Compatibility

A VPN and Tor both alter the normal network path, but they do so differently.

A conventional VPN generally sends covered traffic through one VPN endpoint before forwarding it toward the destination. Real performance depends on factors such as server location, congestion, protocol, route quality, device limits, and the underlying internet connection. If a VPN is unusually slow, VPN slowdown diagnosis works best by comparing the same connection with the VPN off and on before changing several settings at once.

Tor routes traffic through multiple volunteer relays. The Tor Project says Tor Browser can sometimes feel slower because traffic passes through relays in different locations and performance depends on network load and latency. It also notes that users may not always notice a meaningful speed difference.

The distinction matters most for workloads that are sensitive to latency. Interactive voice, video, gaming, and similar applications may react more strongly to additional delay than ordinary page loading, although actual results depend on route and network conditions.

Application coverage also differs. Many VPN clients are designed to route broad device traffic according to their routing policy. Tor Browser is configured for safe web browsing through Tor by default. Core Tor can be used by other compatible applications, but the Tor Project says it has not researched safe anonymity configurations for every program and recommends Tor Browser for normal browsing.

Tor should therefore not be reduced to “a slower VPN.” Its extra relays and application constraints follow from a different privacy architecture.

Should You Use a VPN and Tor Together?

Not by default.

Combining two privacy technologies does not automatically create stronger privacy. It changes which parties occupy different positions in the connection, introduces additional configuration, and can create new failure modes.

The Tor Project’s current Tor Browser known-issues guidance says VPNs tend to interfere with Tor and does not recommend using VPN and Tor together unless the user is advanced and understands how to configure both without compromising privacy.

There is therefore no general rule that stacking the technologies provides “double anonymity.” Whether a combined design changes privacy in a useful way depends on the exact routing arrangement, threat model, VPN operator, and configuration.

This comparison does not provide VPN-over-Tor or Tor-over-VPN configuration instructions. Those designs require a more specific threat model than the general comparison here.

Bottom Line

VPN and Tor overlap in one visible result: either can cause a normal website to see an address other than your ordinary public IP. Their privacy architectures are otherwise substantially different.

A conventional VPN gives you an encrypted path to one VPN endpoint and asks you to trust the operator occupying that position. Tor distributes route knowledge across multiple relays and, when used through Tor Browser, adds browser-level privacy defenses that a VPN tunnel alone does not provide. The useful choice depends on your threat model, application needs, tolerance for latency, and whether concentrating or distributing network trust better matches the problem you are trying to solve.

How Apple’s entry into the foldable market could boost its biggest rival.

The iPhone Duo has sparked more interest in foldable phones

Benjamin Fanjoy/Getty Images

Phone makers should give foldables more attention

Benjamin Fanjoy/Getty Images