WebRTC and DNS: Leaks Nobody Bothers to Close

You flipped on your VPN. Wired up a SOCKS proxy. Funneled DNS-over-HTTPS through Cloudflare. Cracked open the browser. Loaded a site. And that site already knows your true IP, your laptop's name, and the country where your DNS resolver squats. Twelve bucks a month well spent.

Network leaks. Protocol leaks. DNS leaks. Everyone's heard the words. Almost nobody reckons with them. VPN marketing has hawked millions of people on a single Connect button that sweeps away all their privacy woes. Not a scrap of it holds up. What you're buying is a mirage.

Below, we rip apart each specific leak mechanism, backed by academic work from 2022 through 2025. Every single one fires right now, in your browser, regardless of what you've shelled out for privacy.

In brief: WebRTC leaks in every browser except Tor/Mullvad. DoH is fingerprinted at 95% accuracy, ODoH at 94%. mDNS leaks your device name.

WebRTC: A Protocol Nobody Built for Privacy

Web Real-Time Communication, WebRTC, began life as a way to make video calls straight in the browser. No plugins. Google shoved it forward from 2011, Mozilla climbed aboard, Apple bolted it onto Safari in 2017. Today WebRTC hums along in every major browser. And in every major browser, it hemorrhages data.

A team from Sorbonne Université and Snowpack presented their findings at CRiSIS 2024 in November 2024: Preventing WebRTC IP Address Leaks. WebRTC's API was never wrought with privacy in mind. Even after spec updates, sensitive public IPs still spill during audio and video calls. Leaks fire even behind VPN, SOCKS, or HTTP/S proxies. Linux, macOS, Windows. Same wreckage across every platform.

Grasping why demands cracking open ICE: Interactive Connectivity Establishment.

How STUN and TURN Spill Your IP

When two browsers try to stand up a direct WebRTC link, they run ICE to hunt for workable paths. ICE gathers so-called candidates. Three breeds.

Host candidates hold your device's local IP. A familiar 192.168.x.x. Seems harmless. On corporate networks without NAT, and those crop up far more often than folks assume, the host candidate holds the real public IP.

Server-reflexive candidates come from STUN servers. Your browser pings a STUN server; the server fires back: here's your public IP. In certain setups, that request slips past the VPN tunnel altogether. STUN speaks UDP, and not every VPN snags all UDP packets.

Relay candidates flow through TURN servers that ferry traffic on your behalf. TURN should, on paper, mask the real IP. Yet TURN servers demand credentials, and the bare act of hitting a specific TURN server becomes an identifier.

Why HTTP Proxies and SOCKS Fall Short

An HTTP proxy works at the application layer. SOCKS grabs TCP connections. WebRTC speaks UDP. Reaching straight into the operating system's network interfaces, it leapfrogs the proxy stack wholesale. Proxy never sees the traffic. Sails right past like the proxy doesn't exist. Sorbonne's 2024 lab work proved it cold: WebRTC leaks happen even behind VPN, SOCKS, or HTTP/S proxies on every tested platform.

Even a VPN with a kill switch can't guarantee safety here. ICE candidate gathering fires off inside the browser on its own, before the main connection lands. If the VPN client doesn't ensnare every UDP packet at the network driver level, the STUN request bolts free. Most commercial VPNs just don't do that.

ICE Candidates as a Cartograph of Your Network

Here's a separate fingerprinting vector most people miss entirely: the count and types of ICE candidates. A bog-standard home PC with one network interface throws off 2 to 3 candidates. A corporate workstation bristling with several adapters yields 5 to 8. A virtual machine with bridged networking and NAT gives its own telltale set. A laptop running WiFi and Ethernet at once mints candidates for both interfaces. A dual-SIM phone with WiFi awake spawns candidates for three networks simultaneously.

Raw tallies plus type spread sketch a map of your network topology, like a geological core sample. Three host candidates, two srflx, one relay means one setup. One host and zero srflx means another. WHOX check A09 sifts through every ICE candidate and hashes the SDP description whole, yielding a singular signature of your network layout. Any site that asks for a WebRTC connection can read it.

Only one countermeasure actually held in Sorbonne's testing: running Firefox inside a Docker container. Container isolation worked even when the browser itself was compromised. VPN, proxy, SOCKS, all buckled.

October 2025 brought a second study: WebRTC Metadata and IP Leakage in Modern Browsers: A Cross-Platform Measurement Study. Researchers tested 2025 builds of Chrome, Brave, Firefox, and Tor across desktop and mobile. WebRTC leaks remain one of the most underrated privacy holes plaguing VPN users in all major browsers. Worth restating plainly: by 2026, WebRTC is the most critical privacy threat for anyone leaning on VPNs and proxies. Not cookies. Not canvas fingerprinting. Not third-party trackers. A protocol cooked up for video chats turned out to be the most treacherous conduit for leaking real identity. Ships enabled by default in every browser you're running right now.

Comparison Table: What Each Browser Leaks

Data drawn from the Joe12387 matrix, CRiSIS 2024, and arXiv 2510.16168:

Leak	Chrome	Safari	Edge	Firefox	Brave	Opera	Tor / Mullvad
WebRTC IP leak	❌ Vulnerable	❌ Vulnerable	❌ Vulnerable	❌ Vulnerable	❌ Vulnerable	❌ Vulnerable	✅ Disabled
IPv6 leak via VPN	❌ Vulnerable	❌ Vulnerable	❌ Vulnerable	⚠️ Partial	⚠️ Partial	❌ Vulnerable	✅ Protected
QUIC UDP-443 leak	❌ Vulnerable	✅ No QUIC	❌ Vulnerable	❌ Enabled since v88	⚠️ Partial	❌ Vulnerable	✅ Disabled
mDNS device name	⚠️ Partial	⚠️ Partial	⚠️ Partial	⚠️ Partial	⚠️ Partial	⚠️ Partial	✅ Protected
DoH fingerprinting	❌ Vulnerable	❌ Vulnerable	❌ Vulnerable	❌ Vulnerable	❌ Vulnerable	❌ Vulnerable	✅ Via Tor
DoT leak	❌ If enabled	❌ If enabled	❌ If enabled	❌ If enabled	❌ If enabled	❌ If enabled	✅ Protected
DNS geo mismatch	❌ Vulnerable	❌ Vulnerable	❌ Vulnerable	❌ Vulnerable	❌ Vulnerable	❌ Vulnerable	✅ Protected

One column shows green. Just one. Tor Browser and Mullvad Browser, which kill WebRTC outright. Nothing elegant about it. A white flag: the protocol can't be tamed, only gagged. Tor Project devs reached that conclusion after years of trying to rein WebRTC in. Every patch ripped open a fresh leak. Every workaround broke some use case. In the end, the only trustworthy fix was the crudest: media.peerconnection.enabled = false. No WebRTC. No candidates. No leaks. No video calls.

Brave deserves its own note. Branding itself as privacy-first, blocking trackers, randomizing Canvas, Brave ships basic ICE candidate limits and a webrtc.ip_handling_policy toggle. Still doesn't deliver full IP leak protection. Better than stock Chrome. Nowhere near watertight.

DNS: Encryption That Shields Nothing

Use DNS-over-HTTPS and your ISP won't see which sites you visit. You'll hear that refrain across every other tech forum. Technically true. Practically worthless.

DePL: 95% Accuracy from 50 Packets

A system called DePL, Detecting Privacy Leakage in DNS-over-HTTPS Traffic, debuted at IEEE Icnp in 2022. Researchers wielded n-shot learning to sort encrypted DoH traffic.

Result: roughly 95% accuracy in pinpointing which websites a user visits. Through encrypted DNS. No decryption needed.

Parallel work from Loyola University Chicago backed those numbers: 95% accuracy guessing 1 out of 10,000 sites from the first 50 DoH packets. F1-score of 93% in an open-world model spanning 100,000 sites. Yet another paper, Toward Practical Defense Against Traffic Analysis on Encrypted DNS, showed that even with IETF-recommended padding, attackers hit accuracy above 95%.

Why? Dead simple. Encryption hides what's inside packets but leaves their sizes, timings, volumes, and rhythms in plain sight. Every website kicks off a unique chain of DNS lookups: the main domain, CDN subdomains, analytics endpoints, fonts, API calls. Each chain's shape in the traffic is as distinctive as a voiceprint. Encrypted, sure. Still unmistakable.

ODoH: Even a Relay Can't Save You

If DoH falls flat, maybe Oblivious DNS-over-HTTPS will do the trick? ODoH wedges an intermediate relay between you and the DNS resolver. Relay knows your IP but can't read the queries. Resolver reads the queries but doesn't know your IP. A clean architectural split. Cloudflare 1.1.1.1 and Apple iCloud Private Relay both lean on it.

At the IEEE DSN'25 conference, Privacy Analysis of Oblivious DNS over HTTPS: a Website Fingerprinting Study landed. Standing up a test environment on Geni infrastructure, the team threw ensemble deep learning at it.

Result: a passive observer can finger target sites from ODoH traffic at 94% accuracy.

Architectural guarantees crumble under statistics. Double proxying hides who you are from the resolver, yet it can't hide traffic patterns from someone perched on the wire between you and the relay. Your ISP, a government regulator, or the person running the coffee-shop WiFi sees the same packets, the same sizes, the same timing. 94% accuracy.

No chalkboard attack here. A hands-on demo on live infrastructure, published at one of the top conferences on systems dependability.

Leaks You Didn't Know You Had

WebRTC and DNS hog the limelight, yet the roster of protocol-level leaks stretches far longer. Each one alone looks trivial. Woven together, they yield a full profile.

mDNS: Your Laptop Blurts Its Own Name

Multicast DNS was dreamt up for convenience: letting devices on a local network spot each other without a central DNS server. When WebRTC mints ICE candidates, it taps mDNS to craft addresses shaped like .local. Hiding raw IPs was the goal. Succeeded at that. Ripped open a different wound in the process.

Modern browsers tried to patch the local IP leak with mDNS obfuscation. Instead of the real IP, WebRTC spins up a random UUID and registers it as an mDNS name: something like a1b2c3d4-e5f6-7890-abcd-ef1234567890.local. On the other end, mDNS resolves the UUID to the IP, yet the UUID itself betrays nothing. Tidy on paper. In the wild, mDNS requests still broadcast across the local network, and mDNS responses carry the device hostname. UUID shielded the WebRTC candidate from showing the IP through JavaScript. Did not shield mDNS from blurting the hostname at the network layer.

And those mDNS responses contain the device hostname. On a typical MacBook, that reads Ivans-MacBook-Pro.local. On Windows, maybe Desktop-Ivan. Your device name, which routinely holds the owner's real name, gets broadcast across the local network.

In a coffee shop. A coworking space. An airport. A corporate LAN. Anyone on the same subnet watches your machine announce itself: Hi, I'm Ivan's laptop. Random UUIDs in the WebRTC candidate change nothing about that.

Check A14 flags exactly that leak. Data shows most users haven't touched their default device name.

IPv6: Dual Stack Wrecks VPNs

Most VPN clients tunnel IPv4 traffic. IPv6 traffic walks straight out, sidestepping the tunnel completely. If your network speaks both protocols, and more networks do every month, your real IPv6 address lies wide open to any website.

IPv6 leaks cut deeper than IPv4 leaks, and the reason sits at the bedrock of the protocol. An IPv4 address is typically shared: behind a single NAT sit a dozen, a hundred, sometimes a thousand users. IPv6 was engineered to hand every device its own globally unique address. Unlike its IPv4 counterpart, your IPv6 address almost certainly belongs to your exact device. Not your router. Not your subnet. Your laptop, personally. An IPv4 leak gives away a neighborhood. An IPv6 leak gives away a single machine.

Slaac, the protocol for automatic IPv6 address assignment, worsens things. In stock configuration, part of the address gets derived from the network card's MAC address. A unique hardware stamp, welded straight into the IPv6 address. Privacy extensions that randomize that chunk do exist but aren't switched on everywhere by default. Even with privacy extensions, the IPv6 prefix still ties back to your network.

On Windows, Teredo, an automatic IPv4-to-IPv6 bridge, poses a distinct hazard. Stale or misconfigured boxes can dodge the VPN tunnel through Teredo, though most current VPN clients squash it when connected. You believe all traffic flows through the VPN in the Netherlands. Your IPv6 traffic is streaming straight out of Moscow.

Check A11 targets IPv6 leaks through VPNs. A fix exists: turn off IPv6 at the OS level or pick a VPN that handles IPv6 explicitly. Most users do neither.

QUIC: UDP on Port 443 Slips Through

HTTP/3 rides on QUIC, which uses UDP on port 443 rather than TCP. Many VPNs and firewalls only filter TCP. QUIC connections glide right past.

Chrome and the Chromium family lean harder on QUIC whenever the server supports it. Google, YouTube, Cloudflare. Every big platform speaks HTTP/3. Month by month, a growing share of your traffic seeps around the VPN tunnel.

Check A12 catches exactly that. On top of it, check A69 grabs a JA4Q-like fingerprint of the QUIC handshake itself. Even if QUIC doesn't escape the VPN, the QUIC Initial packet pattern alone becomes a fingerprinting vector.

DoT: One More DNS Protocol, One More Leak

DNS-over-TLS runs on port 853. A standalone, easily spotted port. If your system is set up for DoT and your VPN doesn't intercept traffic on 853, your DNS lookups go out raw. Encrypted, yes. Visible to any observer as a distinct stream that's trivially filtered and picked apart.

Just using DoT on port 853 instead of plain DNS on port 53 is itself a marker. Rare setups breed uniqueness. Uniqueness breeds a fingerprint.

Check A22 flags leaking DoT traffic.

DNS Resolver Mismatch: Geography Doesn't Lie

You connected to a VPN server in the Netherlands. Your IP looks Dutch. Your DNS lookups go to Cloudflare in the US. Or, worse, to the DNS server of your Russian ISP.

DNS resolver country mismatch ranks among the most reliable tells for VPN detection. As Proton VPN puts it in their docs: if the IP address of your ISP's DNS server is exposed, that mismatch lets websites figure out you're trying to hide your true identity.

Root cause is plain: modern browsers flip on DoH through their own resolvers by default. Firefox ships DNS lookups to Cloudflare, Chrome to Google DNS. Regardless of VPN settings. Called DoH auto-translation, check A13 nails it.

Check C07 lines up the DNS resolver's geography against the IP address geography. Dutch IP, Russian DNS. Instant red flag. No machine learning needed. Two countries, one comparison, game over.

Cascade Effect: Trifles Compound

Any single leak can seem like noise. A local IP through WebRTC. A device name through mDNS. Where the DNS resolver sits. An IPv6 address. QUIC packet timing.

Now picture the combined profile:

VPN IP address: Germany
DNS resolver: Russia
IPv6 address: France
WebRTC local IP: corporate subnet 10.x.x.x
mDNS hostname: Jean-MacBook
QUIC traffic: bypassing the VPN tunnel
DoH pattern: matches a specific browsing profile at 95%

Seven signals. Seven different protocols and mechanisms. Each by itself could be happenstance. Fused together, they forge a profile sharper than a passport. A person deanonymized. Not because one shield failed. Because none of them work fully, and the accretion of bantam leaks draws an exact portrait.

A crucial wrinkle: these leaks sit at different strata of the network stack. WebRTC lives in the browser. IPv6 lives at the OS level. DNS lives at the resolver. QUIC lives at the transport layer. mDNS lives on the local network. No single tool covers every stratum at once. A VPN works at the network layer but misses WebRTC. Browser toggles close WebRTC but don't govern IPv6. System settings govern IPv6 but don't touch DoH patterns. Each defense mans its own floor of the building. Leaks are scattered across every floor.

Long-Term Observation on Browser Fingerprinting from PoPETs 2020 found the average lifespan of a trackable fingerprint to be 1.8 weeks. Hardware signals linked to GPU, audio, and WebGL hold steady for months and years. Algorithms can stitch fingerprints back together even after they shift, tracing the slow drift: a browser update, a newly loaded font, a screen resolution swap.

What WHOX Checks

WHOX runs 126+ checks across three axes: Fingerprint, Consistency, and Anonymizer. For network leaks, a dedicated cluster stands ready:

Check	What It Catches	Severity
A09	WebRTC leak: host, server-reflexive, and relay candidates; SDP hash	High
A10	DNS resolver country	Medium
A11	IPv6 leak through VPN	High
A12	QUIC UDP-443 leak bypassing VPN tunnel	Medium
A13	DoH auto-translation leak	Medium
A14	mDNS device name leak	Medium
A22	DoT traffic leak	Medium
C07	DNS resolver country vs. IP address country mismatch	High

Check A09 doesn't just ask whether a WebRTC leak exists. Pulling apart every ICE candidate type, host, srflx, and relay, it hashes the session SDP description and cross-checks IP addresses from WebRTC against the declared public IP. Any gap gets logged.

Check C07, DNS geo mismatch, works hand in glove with A10. When the DNS resolver sits in one country and the IP in another, the system records the clash and bumps the A-score. One of the simplest yet most potent signals: faking an IP through a VPN is easy. Getting your DNS lookups to route through that same VPN and the same country turns out to be far thornier.

All of these checks feed into the overall Trust Score at WHOX, where axis F gauges fingerprint uniqueness, axis C gauges signal consistency, and axis A gauges anonymizer detection. Weights: F = 35%, C = 30%, A = 35%. Network leaks hammer two axes at once: A and C.

What to Do About It

Blunt answer: no complete fix exists. Every safeguard plugs some leaks and spawns fresh ones.

Tor Browser kills WebRTC outright. Only browser with a fully green column in the table above. Sluggish, breaks half the web, and the sheer act of using Tor is a signal. Check A01 tags Tor exit nodes as Fatal.

Turning off IPv6 at the OS level seals leak A11. A sound step if your VPN doesn't handle IPv6 on its own. Yet it's a manual tweak that the overwhelming majority of users won't ever bother with.

Turning off QUIC in browser settings seals A12. In Chrome, that's the flag chrome://flags/#enable-quic. Firefox has shipped QUIC by default since version 88 in April 2021 but lets you kill it via about:config with network.http.http3.enable = false. Once more: a manual tweak, a manual re-check after every update.

Swapping your device hostname to something impersonal seals A14. MacBook instead of Ivans-MacBook-Pro. A dead-simple step no one bothers with.

Forcing the VPN's own DNS server instead of the system or browser default seals C07 and partly A13. Browsers keep rolling out their own DoH resolvers that sidestep system settings altogether, though.

Shutting down WebRTC entirely in Firefox via about:config: media.peerconnection.enabled > false. Scorched-earth option. Browser-based video calls stop cold. If you don't use Google Meet or Zoom in the browser, it's the most airtight way to close the whole class of WebRTC leaks in one stroke. One flag, zero candidates, zero leaks. Chrome has no matching built-in switch. Chromium browsers need extensions to block WebRTC, and those extensions themselves become fingerprint signals.

Setting media.peerconnection.ice.default_address_only > true in Firefox. A middle path. WebRTC keeps working, but the browser mints only one ICE candidate instead of the full set. Video calls still function. Local IP leak sealed. Network topology stays hidden. Every Firefox user should flip that switch as a bare minimum.

Sourest truth of all: each step alone plugs one hole. Doing them all, keeping them current, verifying them after every browser and OS update, that's labor only a sliver of people will sustain.

Who Needs to Know

A journalist working with a source in an authoritarian country hooks up to a VPN. WebRTC leaks the local IP of their newsroom network. mDNS broadcasts their laptop's name. DNS lookups flow to a resolver in their real country. A source is now in danger.

From the Knight First Amendment Institute, Silenced by Surveillance landed in February 2025. Grounded in 80+ interviews with journalists and human rights defenders, it lays bare how digital tools have become weapons of transnational repression. Spyware, account breaches, coordinated attacks. Identification sits at the root. And identification starts with a leak.

Activists. Dissidents. Whistleblowers. For them, the gap between almost private and truly private isn't reckoned in abstract percentages. Years behind bars hang in the balance.

Everyday users who bought a VPN for privacy are also exposed. Risk for them isn't prison. Profiling is. Every DNS geo mismatch, every WebRTC candidate, every IPv6 address that squirts past the VPN lands in a real-time ad auction. In January 2025, hackers claimed they'd dumped 17 terabytes of Gravy Analytics data harvested from exactly that auction pipeline. A billion devices. Thousands of apps. Developers of those apps didn't even know their users' data was draining out. FTC had already filed suit against Gravy Analytics and Venntel in December 2024 for collecting and selling sensitive location data. Ad infrastructure hoovers up everything: IP, user agent, geolocation, device IDs. When a VPN lets even one signal through, that signal drops into the same funnel. Not to intelligence agencies. To ad exchanges. Which, as the Gravy Analytics fiasco showed, amount to the same thing.

Conclusions

WebRTC leaks in every browser except Tor and Mullvad, which kill it. VPN, SOCKS, and proxies don't help. Sorbonne/CRiSIS 2024.
DNS-over-HTTPS doesn't deliver privacy. 95% accuracy identifying sites from 50 packets. DePL, IEEE 2022.
Oblivious DoH doesn't deliver privacy either. 94% accuracy even with an intermediate relay. IEEE DSN'25.
mDNS leaks your device name onto the local network. Often it's your real name.
IPv6 bypasses VPNs on dual-stack networks. Most VPN clients don't handle IPv6.
QUIC slips past VPN tunnels on misconfigured systems. Chrome leans on QUIC more every month.
DNS geo mismatch instantly outs VPN usage. Simplest and most telling signal.
Piling up small leaks builds a full profile. Cascade makes anonymity impossible without sealing every hole.

All leaks listed here are caught by checks A09, A10, A11, A12, A13, A14, A22, and C07. If even one check lights up red, your defenses have a breach. Size of the breach doesn't matter. One is enough.

We don't show all of this to strip users bare. We show it because if a system with 300+ signals can assemble your leak profile in seconds, anyone with server infrastructure can do the same. Ad networks already do. State surveillance systems already do. Only question left: do you know about your own leaks before someone else puts them to work.

FAQ

Does a VPN stop WebRTC leaks? No. Sorbonne and Snowpack proved it in the lab at CRiSIS 2024: WebRTC leaks even behind VPN, SOCKS, or HTTP/S proxy on every tested platform. WebRTC speaks UDP and taps OS network interfaces directly, vaulting over the proxy stack. Only fix: kill WebRTC altogether via Tor/Mullvad or set media.peerconnection.enabled = false in Firefox.

What kinds of DNS leaks are there? Six main kinds: 1 DNS resolver mismatch, where DNS sits in one country and the IP in another; 2 DoH fingerprinting, with 95% accuracy identifying visited sites through encrypted DoH per DePL, IEEE 2022; 3 DoT leak, where traffic on port 853 bypasses VPN; 4 DoH auto-translation, where browsers send DNS through their own resolvers sidestepping VPN; 5 ODoH fingerprinting, where even Oblivious DoH is fingerprinted at 94% accuracy per IEEE DSN'25; 6 DNS geo mismatch as a VPN detector.

What is DoH fingerprinting and why doesn't encryption help? Encryption hides the content of DNS lookups but leaves metadata exposed: packet sizes, timings, volumes. Each site kicks off a unique chain of DNS lookups. DePL at IEEE Icnp 2022 hit roughly 95% accuracy from just 50 packets. Loyola University confirmed: 95% prediction of 1 out of 10,000 sites. Even IETF-recommended padding doesn't help. Attackers still reach accuracy above 95%.

What is the mDNS leak and why does it matter? Multicast DNS handles device discovery on local networks. When browsers mint WebRTC ICE candidates, they use mDNS to mask IPs, but mDNS responses carry the device hostname, often the owner's real name like Ivans-MacBook-Pro.local. In a coffee shop, coworking space, or airport, anyone on the subnet sees that name. UUID in the WebRTC candidate changes nothing. Fix: rename your hostname to something impersonal. MacBook instead of your name.

Which browser is safest from network leaks? Only Tor Browser and Mullvad Browser are fully shielded: WebRTC killed, IPv6 blocked, QUIC off, mDNS hushed, DNS routed through Tor or VPN. Every other browser, Chrome, Safari, Firefox, Brave, Opera, is exposed to WebRTC, IPv6, QUIC, and DNS leaks. Firefox lets you close most of them by hand with media.peerconnection.enabled = false, disabling IPv6, and disabling QUIC, but that demands knowledge and steady upkeep after every update.