How to bypass internet censorship in 2026: protocols that actually work

Freedom House's 2025 report says global internet freedom declined for the 15th consecutive year. People in at least 57 of the 72 countries it covers were arrested or imprisoned for online expression, a record high. It calls China and Myanmar the world's worst environments for internet freedom.

These are not abstract statistics. They translate directly into blocked websites, intercepted messages, and people going to prison for running VPN software.

If you need to bypass internet censorship in 2026, the protocol your VPN uses is the only thing that matters. Not the brand name. Not the server count. Not the marketing copy on the landing page. The protocol determines whether your traffic gets through or gets flagged, throttled, and dropped at the ISP level before it ever leaves the country. (What is censorship resistance covers the underlying concept; this post is the concrete protocol-by-protocol assessment.)

This article ranks the protocols that matter by how their design holds up against state-level deep packet inspection. We have not yet run these tests from inside China, Russia or Iran, and we know of no independent test that compares them all, so this is a ranking from design and public research, not from measured success rates. Some of what follows will contradict what you've read elsewhere.

The censorship infrastructure in 2026

The governments doing this are not improvising. They're running purpose-built systems with real engineering teams and real budgets.

China's Great Firewall remains the most advanced censorship apparatus ever built. It combines passive DPI with AI/ML-powered classification and active probing. When the GFW suspects a connection might be a proxy, it doesn't just block it. It connects to the server itself, sends crafted payloads, and tries to provoke a protocol-specific response. If the server answers in a way that a normal web server wouldn't, the IP gets blacklisted. These probes can reach a flagged server soon after the original suspicious connection.

The GFW also does something no other censorship system has replicated at scale: it fingerprints TLS implementations themselves. Your VPN might use TLS 1.3, but if the ClientHello your client generates doesn't match any known browser's fingerprint, the connection dies. The GFW maintains an updated database of browser TLS fingerprints and kills anything that doesn't match.

Russia's TSPU (Technical Systems for Countering Threats) sits inside every licensed ISP in Russia. It's mandatory hardware. Human Rights Watch says the authorities had blocked at least 469 VPNs by February 2026. Mediazona, citing Kommersant, reports plans to expand TSPU capacity at a cost of about $186 million from the federal budget.

TSPU isn't as sophisticated as the GFW. It doesn't do active probing at scale. But it doesn't need to, because most VPN protocols are trivially identifiable through passive analysis alone. WireGuard, OpenVPN, and Shadowsocks are all reported blocked by TSPU.

Iran runs a multi-layer filtering system. A FOCI 2026 paper says QUIC censorship "went into effect before and stayed in effect after" its June 2025 shutdown. Iran uses both tools: outright shutdowns, and targeted, protocol-specific blocking that's hard for journalists to document.

Myanmar has gone further than most. Soldiers have been reported checking civilians' phones at random for VPN applications, and the law can mean prison. The enforcement is physical, not just technical.

What stopped working

If your VPN runs one of these protocols and you're in a censored country, it does not work. Period.

OpenVPN is easy to fingerprint. A 2022 USENIX Security paper identified over 85% of OpenVPN flows with only negligible false positives, and identified connections to 34 of the 41 obfuscated configurations it tested. The control channel framing, the opcode byte structure, the timing of the TLS handshake are all identifiable signatures. Wrappers like Obfsproxy don't reliably hide them.

OpenVPN is fine for privacy on unrestricted networks. For censorship circumvention, it's dead. It has been dead for years.

WireGuard is easy to block. The reason is structural: every WireGuard handshake initiation is exactly 148 bytes with a fixed header. A single conditional in the DPI ruleset catches it. No machine learning needed. UDP traffic, port 51820, first four bytes 0x01 0x00 0x00 0x00, packet length 148. Done.

AmneziaWG 2.0 adds random padding and junk packets, which helps against naive fingerprinting. It doesn't help against statistical analysis of packet size distributions. UDP-based protocols that don't match any known legitimate application are suspicious by default in heavily censored environments.

SSH tunnels have been fingerprinted since at least 2020. The SSH handshake starts with a plaintext version exchange (SSH-2.0-OpenSSH_9.6), which is about as stealthy as announcing your intentions over a loudspeaker.

Basic HTTP/SOCKS proxies don't survive first contact with any DPI system deployed after 2018.

Shadowsocks (pre-2022 versions) suffered from a design flaw that allowed replay attacks. DPI systems would record encrypted handshakes and replay them to the server. The server's response (or lack of response) confirmed it was running Shadowsocks. China exploited this at scale starting in 2020.

Protocols ranked by DPI resistance

We rank these by how their design holds up against the three hardest censorship systems: China's GFW, Russia's TSPU, and Iran's filtering infrastructure. We publish no success or detection rates here. We found no independent measurement of these protocols by country, and we have not yet run our own test from inside these countries.

1. VLESS Reality with the Vision flow

VLESS Reality with the Vision flow (xtls-rprx-vision) is the best anti-censorship protocol available in 2026. In our assessment, nothing else fits advanced DPI as well by design.

The reason is architectural. Reality doesn't try to obfuscate VPN traffic. It doesn't add padding or inject noise. Instead, it makes VPN connections produce TLS handshakes that are byte-for-byte identical to connections to a real website. When you connect to a VLESS Reality server, the TLS negotiation your client performs is indistinguishable from a genuine connection to, say, www.microsoft.com. The DPI system sees a normal TLS 1.3 handshake with a normal SNI, normal cipher suites, and a certificate chain that belongs to a real website.

This works because Reality uses the camouflage target's actual certificate during the TLS handshake, then performs an authenticated key exchange hidden inside the TLS session. The DPI system would need to MitM the connection to detect it, which breaks TLS entirely and is impractical at national scale.

The Vision flow (xtls-rprx-vision) added in late 2023 eliminates the TLS-in-TLS detection signal that traffic analysis would otherwise produce. That signal is a real risk: a 2024 USENIX Security paper shows encapsulated TLS handshakes can be detected at an ISP. It doesn't name Vision, and we found no independent test that settles how Vision fares.

During known crackdown periods (around Party Congress sessions, Tiananmen anniversary), the GFW is reported to tighten its heuristics and may block IPs with unusually high traffic volumes, regardless of protocol. That's IP reputation, not protocol detection.

2. Hysteria 2

Hysteria 2 takes a different approach. It's built on QUIC (UDP-based, like WireGuard) but adds two features designed for hostile networks. "Brutal" congestion control ignores packet loss signals and maintains bandwidth even when the network is actively interfering with the connection. "Salamander" obfuscation modifies QUIC headers to resist fingerprinting.

The protocol is fast. On networks where it works, Hysteria 2 often outperforms VLESS Reality on throughput because Brutal congestion control doesn't back off when it detects loss. This is useful in environments where the censor is throttling rather than outright blocking.

The problem is that QUIC itself is increasingly scrutinized. China's GFW has been experimenting with blocking all non-standard QUIC traffic, and Hysteria 2 connections have been caught in those sweeps. The protocol works well in Russia and Iran today. Its long-term viability in China is uncertain. Being UDP-based is a liability in countries where the censor can simply drop all UDP traffic that doesn't match known services.

3. Trojan-GFW

Trojan takes the "look like real HTTPS" approach, similar in spirit to VLESS Reality but with a weaker implementation. The Trojan server runs a real web server (typically Nginx) behind TLS. Trojan traffic goes to the proxy handler; non-Trojan traffic gets served the real website. From the outside, the server genuinely is running a website.

The fatal weakness is active probing. When the GFW or TSPU connects to a suspected Trojan server and sends it non-Trojan traffic, the server responds with a web page. That's expected. But the server's behavior in edge cases (connection handling, error responses, timeout patterns) differs subtly from a genuine Nginx deployment. These differences can give a Trojan server away.

Trojan was a good protocol in 2021. It's showing its age. If you're running it and it still works for you, great. Don't set up new infrastructure on it.

4. TUIC v5

TUIC multiplexes QUIC streams and adds its own authentication layer. It's newer than Hysteria 2 and has seen less deployment. Like Hysteria 2, it inherits QUIC's vulnerability to blanket UDP filtering. The smaller user base means less testing against production DPI systems and less certainty about detection rates.

Potentially viable, but not battle-tested enough to recommend as a primary protocol.

5. Shadowsocks 2022 (AEAD-2022)

The 2022 revision of Shadowsocks fixed the replay attack vulnerability and added proper AEAD encryption with key derivation per session. This was necessary and overdue.

But Shadowsocks still has the entropy problem. The encrypted stream has uniformly high entropy from the first byte. Normal HTTPS connections have structured elements (handshake, certificate chains, HTTP framing) that produce recognizable entropy patterns. A 2023 USENIX Security paper found the Great Firewall blocking fully encrypted traffic in real time using simple rules on the first bytes, on about a quarter of connections to the server ranges it watches.

Shadowsocks 2022 is better than its predecessors. It's still fighting with one hand tied behind its back compared to protocols that produce traffic indistinguishable from legitimate HTTPS.

6. Commercial VPN proprietary protocols

NordVPN launched NordWhisper in January 2025, describing it as "web tunnel technology" that disguises traffic as HTTPS. Designed for Wi-Fi restrictions and corporate firewalls. ProtonVPN offers Stealth, which wraps WireGuard in an obfuscation layer.

Both are closed-source. Neither has been tested against the GFW or TSPU by independent researchers. NordWhisper was designed for hotel WiFi and corporate networks, not nation-state DPI. That's a different threat model entirely. ProtonVPN Stealth wrapping WireGuard doesn't change the fundamental problem that the underlying protocol's traffic patterns leak through obfuscation layers over time.

Closed-source protocols can't be audited by the censorship circumvention research community. When a team at the University of Michigan or the Open Technology Fund tests anti-censorship tools, they test open-source implementations. NordWhisper and Stealth don't get that scrutiny. You're trusting marketing claims without independent verification.

7. Tor with pluggable transports

Tor still works via obfs4 bridges and Snowflake. It's slow. For accessing a specific blocked website or sending a message, Tor with bridges is viable. For anything resembling normal internet usage (video, large downloads, real-time communication), it's not practical.

Tor's strength is the network design (onion routing provides anonymity that no VPN can match), not the transport layer. If your threat model requires anonymity from the VPN provider itself, Tor is the only option. If your threat model is "access the open internet from a censored network," faster tools exist.

Why one protocol isn't enough

Here's the part that most VPN providers don't want to talk about: censorship systems adapt.

What gets through today may not get through a year from now. The GFW team reads the same research papers we do. They attend the same conferences. They know how Reality works. They're working on detection methods.

When a single protocol fails, a single-protocol VPN becomes a brick. You're disconnected with no fallback.

The correct architecture is protocol rotation. Your client should try the fastest protocol first. If it fails or gets blocked, it should automatically fall back to a more resistant protocol without manual intervention. The user shouldn't need to know what VLESS is or what DPI means. They press connect and the software figures out what works on their network.

Fexyn runs three protocols with automatic rotation: Bolt (WireGuard) for speed on unrestricted networks, Stealth (VLESS Reality with Vision flow) for censored networks, and Secure (OpenVPN) as a compatibility fallback. The rotation engine classifies failures in real time. If WireGuard gets blocked, the next attempt goes through VLESS Reality+Vision over TLS 1.3. If Stealth fails, OpenVPN over TCP port 443 is the last resort.

Two strategies ship by default. SpeedFirst tries WireGuard, then VLESS Reality+Vision, then OpenVPN. StealthFirst reverses the order, starting with Reality+Vision for users who know they're on hostile networks. The kill switch stays active through the entire rotation cycle. Your real IP never leaks during protocol switching.

No single protocol will work everywhere forever. The censorship arms race guarantees that. What works is having multiple protocols and switching between them faster than the censor can adapt.

Practical advice by country

What follows is based on public reports, not on our own testing inside these countries. Conditions change. What works in March may not work in June. Treat this as a starting point, not gospel.

China

Use VLESS Reality with the Vision flow. Its design best fits what the GFW is known to do. Connect to servers outside mainland China; from Fexyn's fleet, Cyprus is closest geographically and Frankfurt or Ashburn are alternatives if Cyprus is congested.

Expect degradation during politically sensitive periods, such as annual legislative sessions, anniversary dates, and any period of domestic unrest. Having a backup server in a different country on a different IP range helps.

Hysteria 2 (not shipped by Fexyn; relevant for self-hosters) works intermittently and isn't reliable enough to be a primary choice.

Do not use WireGuard, OpenVPN, or Shadowsocks as your primary protocol. They are blocked.

Russia

VLESS Reality works today. TSPU is the primary adversary, and it doesn't do active probing at the GFW's level. Reality's camouflage holds up well against passive DPI and ML classifiers.

The risk in Russia is the pace of TSPU development. Funding keeps growing, and so does the count of blocked VPNs. The system is getting better fast.

Server choice matters. Use IPs that haven't been flagged. Shared hosting IPs (where your VPN server shares an IP with legitimate websites) are harder for TSPU to block without collateral damage. Dedicated VPN server IPs on well-known cloud providers (AWS, DigitalOcean, Vultr) get blocked more aggressively because blocking them causes less collateral damage.

Iran

VLESS Reality with the Vision flow is the protocol class that survives. Iran's DPI technology is less sophisticated than China's or Russia's, but the government compensates with periodic total shutdowns and aggressive throttling.

Iranian ISPs sometimes throttle all encrypted traffic to unusable speeds rather than blocking it outright. For users seeing throttling rather than outright blocking, switching server location (Cyprus, Frankfurt, Helsinki, Ashburn) often helps more than switching protocol.

Myanmar, Turkmenistan, North Korea

These countries combine technical censorship with physical enforcement. Myanmar's law can mean prison for VPN use. Turkmenistan forces citizens to swear oaths not to use VPNs. North Korea sends people to labor camps.

Technical circumvention advice is secondary to personal safety. If you're in one of these countries and need to access blocked content, the technical protocol barely matters compared to the operational security around how you obtain, install, and use the software. Don't install VPN apps on a device that could be inspected at a checkpoint. Don't discuss circumvention tools on monitored communication channels. The protocol is the easy part. Not getting caught is the hard part.

Legal risks are real

Bypassing censorship is illegal in many countries. The penalties vary widely.

Country	Penalty for VPN use
China	Prison possible (rarely enforced against individuals; enforcement targets VPN sellers and operators)
Russia	Fines for companies advertising VPNs; individual use is in a legal gray area but enforcement is increasing
Iran	Prison and fines possible
UAE	Large fines and prison for using VPNs for illegal activity
Myanmar	Prison possible
North Korea	Labor camps
Turkmenistan	Fines and prison; citizens have been forced to swear oaths on the Quran not to use VPNs

We are not lawyers. This table is not legal advice. Laws change. Enforcement patterns change. Research the current legal situation in your specific country before using any circumvention tool.

The practical reality is that in most of these countries, many people use VPNs daily and enforcement against individual users is selective rather than universal. In China, prosecutions of individual users are rare and typically come with other charges. Russia primarily targets VPN providers and advertisers, not end users. But "selective enforcement" also means enforcement can be directed at anyone the government decides to target, for any reason.

Country-specific guides

Each of the countries above has its own threat model, regulator, and recommended protocol. The deeper per-country guides cover the carrier-level details:

• VPN for India: post-CERT-In honest framing
• VPN for Pakistan: surviving the December 2025 PTA crackdown
• VPN for Turkey: BTK throttling and Stealth
• VPN for Russia: protocols that survive TSPU
• VPN for Belarus: Russian-language first, severe enforcement
• VPN for UAE: VoIP unblock, expat market
• VPN for Saudi Arabia: CITC filtering, Hajj/Umrah angle
• VPN for Egypt: install before you fly
• VPN for Vietnam: Telegram unblock since June 2025
• VPN for Kazakhstan: surviving the Qaznet MitM threat model
• VPN for Indonesia: Reddit/Vimeo/Steam unblock
• VPN for Nigeria: the Twitter-ban playbook
• VPN for Ukraine: wartime privacy, occupied vs controlled

The app-specific guides cover the platforms the censors target most:

• VPN for WhatsApp: Gulf and Egyptian VoIP blocks
• VPN for Telegram: Vietnam, Iran, Russia restrictions
• VPN for Twitter / X: Nigeria, Turkey, Pakistan, Russia

For the cross-country picture, the global censorship map is an interactive visualisation of the same markets: filter by VoIP-blocked, DPI-deployed, or VPN-restricted to see the patterns at a glance.

What comes next

The censorship arms race will continue. It always has. Governments spend more money on DPI. Protocol designers find new ways to make VPN traffic indistinguishable from legitimate traffic. The cycle repeats.

Three trends will shape the next two years.

First, encrypted client hello (ECH) is rolling out across major browsers. ECH encrypts the SNI field in TLS handshakes, which currently leaks which website you're connecting to. Widespread ECH adoption would make TLS-based censorship much harder, because the censor can't see the domain name in the handshake anymore. It would also change the camouflage calculus for protocols like VLESS Reality, which currently rely on the censor being able to see (and trust) the SNI.

Second, AI-powered traffic classification is getting better. Russia's investment in ML-based filtering is not unique. Every country with a DPI program is experimenting with neural network classifiers that detect VPN traffic based on behavioral patterns rather than protocol signatures. These classifiers produce more false positives than signature-based detection, which limits how aggressively they can be deployed. But the accuracy is improving year over year.

Third, collateral damage is the main constraint on censors. Blocking an IP range that hosts both VPN servers and legitimate businesses has economic costs. Protocol designers exploit this by making VPN traffic look identical to traffic that the censor can't afford to block. VLESS Reality impersonating microsoft.com works because China isn't going to block microsoft.com. If they did, a lot of the enterprise software in the country would break.

The protocols that survive will be the ones that raise the cost of detection above the cost the censor is willing to pay. That's been true since the first firewall was deployed, and it'll be true in 2030.

Fexyn ships VLESS Reality with the Vision flow, WireGuard, and OpenVPN with automatic protocol rotation. If you're in a censored country, the StealthFirst rotation strategy starts with Reality+Vision and falls back automatically. The use-case framing is on VPN for journalists and VPN for activists; for a head-to-head with a privacy-first competitor that doesn't ship Reality, see Fexyn vs Mullvad. Download Fexyn and try it on your network. All three protocols are included in every plan. The protocol rotation handles the rest.