REALITY, Xray, and AmneziaWG: The 2026 Anti-Censorship Stack
If you are shipping a VPN into China, Iran, or Russia, vanilla WireGuard and OpenVPN will not survive the network. Deep packet inspection fingerprints them quickly and the connection dies. So the real question for anyone building a VPN product for censored markets is not “which protocol is fastest.” It is “which protocol still gets through, and for how long.”
Three protocols carry most anti-censorship traffic in 2026: REALITY, Xray with VLESS, and AmneziaWG. This explains what each one does, how it hides from DPI, where it falls down, and which combination is worth shipping. It is written for the people building the apps and infrastructure, not for end users picking a VPN to install.
One honest caveat first: none of this is a guarantee. Whether traffic gets through depends on the user’s ISP, the filtering on their network, your server IP reputation, the DPI rules in force that day, and how carefully the obfuscation is configured. Treat censorship resistance as ongoing work, not a feature you ship once.
How REALITY hides VLESS inside a real TLS handshake
REALITY is a transport obfuscation layer that makes proxy traffic look like an ordinary visit to a major website. When a client connects, the REALITY server sits in the middle of a genuine TLS handshake with a real public site, for example www.apple.com or www.microsoft.com, and an authorized client negotiates a hidden tunnel on top of that handshake. To anything watching the wire, it looks like a normal HTTPS connection to a site no censor wants to block.
The clever part is what happens under probing. If a censor repeats the handshake to investigate the server, REALITY falls back to serving the real certificate of the cover site, so the server looks identical to the legitimate destination. Blocking it then means blocking Apple or Microsoft, which carries a political and economic cost most censors will not pay.
REALITY is one of the stronger stealth options in 2026, but it is not free. It is more complex to deploy and maintain than a plain tunnel, and its effectiveness still depends on choosing a good cover target and keeping your server IPs clean. A sloppy REALITY setup is detectable.
How REALITY still gets detected
Detection usually comes from active probing, specifically SNI fuzzing. A censor connects with the real server name, then connects again with a small change such as a made-up subdomain. A real CDN answers both the same way. A REALITY server often behaves slightly differently on the altered probe, because it is delegating to its fallback instead of serving the request natively. The tell might be a different cipher order, a different session ticket length, or a small timing difference.
The probe is imperfect and does not catch a well-configured server every time, but it is reliable enough to move a verdict from clean to suspicious when other signals agree. The practical lesson for builders: target selection and configuration discipline are the difference between REALITY that survives and REALITY that gets flagged.
Xray and VLESS: the transport everything wraps
Xray is the actively developed fork of V2Ray, and it is where most serious REALITY deployments live. It implements VLESS, a lightweight stateless proxy protocol designed to replace VMess. VLESS deliberately carries almost no obfuscation of its own. That is the point: it decouples the transport from the disguise, so you can wrap it in REALITY, WebSocket, gRPC, or XHTTP and switch wrappers as conditions change.
A common 2026 deployment layers REALITY over VLESS on a small VPS running an unmodified nginx or caddy as the cover service, often behind Cloudflare for an extra layer of blending. Self-hosted REALITY pointed at a Cloudflare edge has one of the higher observed survival rates inside China, Russia, and Iran, though “observed” is doing real work in that sentence. Results vary by region, ISP, and timing.
If you are weighing Xray against other engines for a multi-protocol client, our Xray vs sing-box comparison covers the trade-offs.
AmneziaWG: obfuscated WireGuard you can hand to non-technical users
Vanilla WireGuard is easy to fingerprint. Its handshake-initiation packet has a fixed structure and size, and it defaults to UDP/51820. Both are trivial for DPI to match. AmneziaWG, the obfuscated WireGuard fork from the Amnezia team, changes that on three fronts:
- Packet shape randomization. The handshake packet gets random-length junk added, so size-based signatures miss it.
- A junk-byte warm-up. Before the real handshake, the client sends several random-sized junk packets to mask the handshake timing pattern.
- Configurable ports. It commonly runs on a random high UDP port or TCP/443 instead of UDP/51820.
The reason AmneziaWG matters for product teams is reach. It is the default in AmneziaVPN’s client apps across iOS, Android, Windows, macOS, and Linux, which makes it the most approachable obfuscated WireGuard option for non-technical users. It keeps WireGuard’s speed while adding enough disguise to get through Russia’s TSPU and Iran’s filtering, and it has worked against China’s Great Firewall as well.
That last point is the volatile one. The Great Firewall keeps upgrading its detection, and junk-byte patterns are exactly the kind of thing machine-learning classifiers learn to spot, so do not assume today’s config survives next quarter. For how AmneziaWG sits next to WireGuard, OpenVPN, and the rest, see our comparison of VPN protocols.
Which combination to ship in 2026
For the hardest environments, the stack most builders converge on is Xray VLESS plus REALITY, pointed at a Cloudflare edge IP, hosted on a small VPS outside the censored country, with AmneziaWG as the fallback when REALITY is disrupted.
The logic is operational cost, not magic. REALITY traffic to Cloudflare is hard to separate from the millions of legitimate connections to Cloudflare, and Cloudflare is too large to blanket-block. When REALITY gets disrupted, usually during politically sensitive windows, AmneziaWG on random UDP ports can carry users until the DPI systems retune. Two transports with different failure modes beat one.
Here is the short version, with the limitation that matters most for each:
| Protocol | Best for | Main limitation |
|---|---|---|
| Xray VLESS + REALITY | Stealth in the strictest networks | Complex to deploy; depends on cover target and IP reputation |
| AmneziaWG | Approachable obfuscated WireGuard, strong fallback | Junk-byte patterns are a moving target as DPI improves |
| Vanilla WireGuard / OpenVPN | Mainstream markets and speed | Fingerprinted and blocked in heavily censored networks |
Whether any of this gets through depends on the user’s ISP, their network’s filtering, the reputation of your server IPs, the DPI rules in force at the time, and how the protocols are configured. Plan for continuous testing and multiple fallbacks, not a one-time setup.
Build with detection in mind
It helps to know how your exit servers look from the other side, because that is where a lot of VPN products quietly fail. A clean disguise on the client connection does not matter if the exit IP announces “datacenter VPN.”
A few things consistently give exit servers away:
- Datacenter IP ranges. REALITY and AmneziaWG exits on Hetzner, OVH, or DigitalOcean are classifiable as datacenter IPs with high confidence. That alone is a strong signal to detection systems and to the streaming and banking sites your users care about.
- Default configs and predictable ports. A random UDP port is a weak signal on its own, but combined with a probeable endpoint it adds up.
- Answering probes. A WireGuard or AmneziaWG endpoint that has not been configured to drop unrecognized initiators will answer a handshake probe, which confirms what it is.
The build-side takeaway is concrete. Invest in IP reputation and sourcing, use residential or clean ranges where it matters, configure endpoints to stay quiet to unauthenticated probes, rotate addresses that get flagged, and test per region rather than assuming a working config is permanent. This is the unglamorous operational work that decides whether a censored-market VPN actually works for real users.
FAQ
Is AmneziaWG enough on its own to get through the Great Firewall?
Sometimes, and not reliably. AmneziaWG has worked against the Great Firewall, but China’s detection keeps improving and junk-byte obfuscation is the kind of pattern machine-learning systems are getting better at flagging. For the strictest networks, pair it with a REALITY-based stack and a fallback, and test continuously rather than trusting a single protocol.
REALITY or Shadowsocks-2022 for China?
Both are in current use and neither is a guarantee. REALITY’s strength is camouflaging as a major site and resisting active probing. Shadowsocks-2022 hardened the older Shadowsocks against probing and is lighter to run. Many deployments use REALITY as the primary stealth layer and keep another transport in reserve. The right answer depends on your IP reputation, your cover targets, and per-region testing.
Can I just ship vanilla WireGuard for users in Iran or Russia?
No, not on its own. Vanilla WireGuard is fingerprintable and is regularly blocked by Russia’s TSPU and Iran’s filtering. You need an obfuscation layer such as AmneziaWG, or WireGuard tunneled inside another transport, plus fallbacks.
Do these protocols work inside mobile apps?
Yes, with platform work. On iOS the tunnel runs in a Network Extension Packet Tunnel Provider with a tight memory budget. On Android it builds on the VpnService API with a foreground service and user consent. Multi-protocol clients are usually built on an engine like sing-box so the app can fall back between transports. Reliability across devices and OS versions is real engineering, not configuration.
The short version
In 2026 the working anti-censorship stack for a serious VPN product is Xray VLESS with REALITY as the primary stealth layer and AmneziaWG as the fallback, hosted on clean IPs outside the censored country. But the protocols are only half of it. The other half is IP sourcing, endpoint hygiene, and per-region testing that never really stops. Ship the stack, then budget for keeping it alive. If you are scoping the whole product around this, our guide on how to start a VPN business in 2026 covers the rest.
Targeting a restrictive market? digitalD.tech can help you scope a protocol and obfuscation strategy, the server and IP setup behind it, and a per-region testing plan to keep it working. Tell us where your users are and we will tell you what we would build. Talk to us, or see our white label VPN options.