XHTTP in China and Iran: why one protocol behaves so differently
XHTTP in China and Iran: a week of HopFox VPN data, compared with Japan, Hong Kong and Germany, and what is censorship versus infrastructure.
According to HopFox VPN, during the week of October 30 to November 5, 2026, the XHTTP transport connected in 89% of 19,423 attempts in China and in 62% of 27,141 attempts in Iran. The Chinese figure matches Japan and Hong Kong, where there is no heavy censorship. The Iranian figure is 27-29 points below them, but only 4 points below Germany, which has no censorship at all. So the gap between China and Iran is made of at least two parts: what happens to international connectivity in Iran, and how well a particular server network reaches a particular region. Which part is bigger cannot be settled from one week of one network. But the failure pattern in Iran differs noticeably from China’s, and that is worth unpacking.
What XHTTP is and how it differs from Reality
XHTTP is a transport in Xray-core, one of the main open source censorship circumvention tools. In mid-2024 Xray gained a mode called SplitHTTP, which slices tunnel traffic into ordinary HTTP requests and responses. Late that same year the mode was reworked and renamed XHTTP.
A classic VPN holds one long connection with a recognizable packet pattern. XHTTP looks like a browser downloading a file: data to the client arrives as a long server response, data from the client leaves as a series of POST requests or a separate stream. It runs over HTTP/1.1, HTTP/2 and HTTP/3, so it can be routed through a CDN such as Cloudflare. The censor then sees a connection to a CDN address, not to a VPN server.
Reality solves a different problem: it hides the server itself. It was added to Xray in 2023. The client starts a TLS handshake as if visiting a real, large website. If the server recognizes its own client by a secret key, the tunnel opens. If a stranger knocks, for example a censor’s probe, the server passes it on to the real website, and the prober sees a genuine certificate.
They are layers, not rivals: XHTTP is often wrapped in Reality or in plain TLS. In HopFox VPN statistics they appear as separate rows, two ways to connect, both trying to look like ordinary HTTPS.
Why the old VPN protocols stopped working
Every generation of circumvention lived until the censor found its fingerprint.
| Protocol | What gives it away | Status in China, Iran and Russia |
|---|---|---|
| OpenVPN | recognizable handshake start | cut by signature |
| WireGuard | fixed bytes in packet headers | cut by signature |
| Shadowsocks, VMess | traffic looks like structureless noise | in China the GFW catches such traffic by byte randomness, a mechanism described by GFW Report researchers and co-authors at USENIX Security 2023 |
| AmneziaWG | masks WireGuard headers but runs on UDP | works where the network does not cut UDP |
| Reality | looks like HTTPS to a real website | works, figures below |
| XHTTP | looks like ordinary web traffic, can go through a CDN | works, figures below |
A detailed country map, including Russia, is in the KRAKOBIT article VPN in China, Iran and Russia: what is blocked and what works.
How many XHTTP connections get through in China and Iran
Here and below: the share of successful connections on the HopFox VPN network over seven days, counted across all attempts. Country means the country of the network the client connected from. The gap is China minus Iran, in percentage points.
| Protocol | China | Iran | Gap |
|---|---|---|---|
| Reality | 91% of 27,699 | 71% of 21,027 | 20 points |
| XHTTP | 89% of 19,423 | 62% of 27,141 | 27 points |
| AmneziaWG | 34% of 2,376 | 32% of 2,666 | 2 points |
China leads Iran by 20 points on Reality and by 27 on XHTTP. On AmneziaWG the two are nearly equal at about a third of connections. That is a story about UDP, covered below.
The samples are comparable. Across the three protocols, Iran has about 50,800 attempts and China about 49,500.
Why China looks like Japan and Germany does not
A figure of 62% or 89% says little on its own. It has to be read next to control countries without heavy censorship: Japan, Hong Kong, Germany and the United States. If those fail too, the cause is servers, routes or the app, not the censor.
| Country | Reality | XHTTP |
|---|---|---|
| Japan | 92% of 1,756 | 91% of 1,355 |
| Hong Kong | 91% of 2,037 | 89% of 1,694 |
| Germany | 76% of 2,259 | 66% of 2,948 |
| United States | 72% of 520 | no attempts |
| China | 91% of 27,699 | 89% of 19,423 |
| Iran | 71% of 21,027 | 62% of 27,141 |
Now the same picture as differences. Each cell is the censored country minus the control country, in points. Minus means the censored country does worse, plus means better.
| Control country | China, Reality | China, XHTTP | Iran, Reality | Iran, XHTTP |
|---|---|---|---|---|
| Japan | -1 | -2 | -21 | -29 |
| Hong Kong | 0 | 0 | -20 | -27 |
| Germany | +15 | +23 | -5 | -4 |
| United States | +19 | no data | -1 | no data |
Against Japan and Hong Kong, China differs by 0-2 points, which is within noise. Against Germany and the US, China is 15-23 points higher. Iran is 20-29 points below Japan and Hong Kong, but only 1-5 points below Germany and the US.
The careful conclusion: on this network, connection success depends heavily on region, not only on censorship. Clients in Germany and the US connect noticeably worse than Japanese ones, even though nobody filters them. Why is not visible from these figures: it could be server placement, route quality, the kinds of networks clients use. By level, China sits next to Japan and Hong Kong, Iran next to Germany and the US.
The other uncensored countries fall between these poles. Singapore (91% on Reality, 87% on XHTTP), France (90% and 88%) and the United Kingdom (84% and 82%) are closer to the Asian level, while the Netherlands shows 70% on Reality. Samples there range from 438 to 1,199 attempts.
How China and Iran block the internet
Independently of these figures, here is what is known about the two systems.
China’s GFW primarily identifies protocols. In 2020 GFW Report researchers described how it sends active probes to suspected Shadowsocks servers and blocks those that respond in a telltale way. In 2023 the same group and co-authors showed how the GFW catches fully encrypted traffic that looks like random noise. Reality and XHTTP were designed as an answer to exactly these two methods: a probe sees a real website, and the traffic looks like ordinary HTTPS rather than noise.
Iran is building the National Information Network (NIN), a domestic network for state and local services. The point is that the outside world can be cut off without switching off the country. In November 2019, during protests over fuel prices, Iran cut international internet access for almost a week while some domestic services kept working.
This suggests a hypothesis that the data can support or not. If China’s filter mainly identifies protocols while Iran’s system more often hits connectivity to the outside world as a whole, then a stealthy protocol in China should behave as it does in uncensored neighbors. In Iran, failures should be dominated by cases where the connection opened but no answer came back, and cases where nothing answers at all. The failure pattern below is consistent with this. It is not proof.
Why the tunnel in Iran opens and then goes silent
The most frequent failure reasons over the week, number of cases:
| Failure reason | Iran | China |
|---|---|---|
| No answer through the tunnel | 10,596 | 3,044 |
| No answer through the tunnel on the HTTP check | 2,074 | not among frequent |
| No answer from any location on the client’s network | 2,588 | not among frequent |
| No AmneziaWG handshake, UDP blocked | 1,309 | 1,307 |
Raw counts mislead: Iran simply has more failures. Converting success rates into failure counts gives about 18,000 failures in Iran over the week and about 6,000 in China (an estimate based on rounded percentages). That is roughly 3 to 1. So the 3.5x gap in silent tunnels mostly reflects the overall failure level, not a uniquely Iranian mechanism.
Shares within failures are more telling (all figures here are estimates):
| Share of all failures | Iran | China | Azerbaijan | Russia |
|---|---|---|---|---|
| Silent tunnel | about 70% | about 49% | about 71% | about 67% |
| No location answers | about 14% | not among frequent | not among frequent | not among frequent |
| UDP blocked for AmneziaWG | about 7% | about 21% | no data | no data |
For Azerbaijan (1,075 silent tunnels) and Russia (893), the calculation uses only Reality and XHTTP, since no other protocols appear in their statistics.
The silent tunnel turns out to be the main failure reason not only in Iran but also in Russia and Azerbaijan. In China its share is lower, partly because AmneziaWG with blocked UDP takes a larger slice of failures there.
The tunnel opened but stays silent
This error means the connection to the server was established, but a test request through it never came back. There can be several causes: loss and throttling on the international link, a filter dropping the connection after the handshake, problems on the server side. A counter cannot tell them apart, and which protocol hits this error more often is not visible either.
XHTTP builds its tunnel from many HTTP exchanges, so on a lossy link it should, in theory, have a harder time fitting into a timeout than Reality. This is a hypothesis, not a measurement.
No location answers
2,588 times the app in Iran could not reach a single server of the service. The protocol is irrelevant here. Most often this is what a network looks like when the operator has closed the exit abroad or lets traffic through only by allowlist, though it can also be a poor connection on the client’s side. In China this reason is not among the frequent ones.
UDP is cut in both countries
The AmneziaWG handshake failed 1,309 times in Iran and 1,307 times in China. That is about 72% and 83% of all AmneziaWG failures respectively (estimate). AmneziaWG hides the WireGuard fingerprint well but runs on UDP, which many networks in both countries cut or throttle. More detail is in the article AmneziaWG in Iran and China: why it connects only now and then. For Iran there are also approaches that need no VPN servers at all: KRAKOBIT covered Shroudly, which bypasses filtering without a VPN, account or servers.
Which is better in Iran: Reality or XHTTP
Over the week in Iran, Reality succeeded in 71% of cases and XHTTP in 62%, a 9-point gap in Reality’s favor. In China the same pair gives 91% and 89%, a 2-point gap.
These figures cannot be compared head to head inside Iran. There the server offers XHTTP first, and after a failure the app switches to another protocol on its own. XHTTP takes the first attempts, including during the network’s worst moments. Reality often gets the second attempt, sometimes after the network has recovered. Only one thing can be claimed: on this network, both protocols are alive in both countries.
Did things get worse during the week
For each country and protocol pair there is a success rate at the start and at the end of the week. The Change column is based only on these: end minus start, in points. The weekly figure in the tables above is counted across all attempts at once, so it lies between the start and end values and leans toward the part of the week with more attempts. Example: Iran, Reality, 74% at the start, 68% at the end, 71% for the week.
| Country and protocol | Start | End | Change |
|---|---|---|---|
| China, Reality | 93% | 91% | -2 |
| China, XHTTP | 91% | 89% | -2 |
| China, AmneziaWG | 34% | 34% | 0 |
| Iran, Reality | 74% | 68% | -6 |
| Iran, XHTTP | 64% | 61% | -3 |
| Iran, AmneziaWG | 34% | 31% | -3 |
| Japan, Reality | 90% | 93% | +3 |
| Japan, XHTTP | 90% | 91% | +1 |
| Hong Kong, Reality | 95% | 90% | -5 |
| Hong Kong, XHTTP | 88% | 89% | +1 |
| Germany, Reality | 75% | 76% | +1 |
| Germany, XHTTP | 67% | 64% | -3 |
| United States, Reality | 73% | 70% | -3 |
In the four control countries over the same week, changes fall within a corridor from -5 (Hong Kong, Reality) to +3 (Japan, Reality). China, plus Iran on XHTTP and AmneziaWG, fit inside this corridor. Only Iran on Reality falls outside it: -6, one point below the lower bound. One week is not enough to speak of a new wave of blocking, but this row is worth watching.
Countries with small samples swing harder: the Netherlands, Reality, -18 (438 attempts in the week), France, XHTTP, +13 (765), the United Kingdom, XHTTP, +9 (754), Singapore, XHTTP, +8 (1,094). Censorship does not explain such swings. This is the behavior of servers, routes and small numbers.
How XHTTP works in Russia and Azerbaijan
| Country | Reality | XHTTP | Reality change | XHTTP change |
|---|---|---|---|---|
| Russia | 66% of 2,014 | 74% of 2,464 | +2 | +2 |
| Azerbaijan | 76% of 2,215 | 69% of 3,182 | -1 | +1 |
Russia is the only censored country where XHTTP beats Reality on this network, by 8 points. In Azerbaijan it is the reverse: Reality leads by 7. By level, both countries sit closer to Iran, Germany and the US than to China. Samples here are roughly 10 times smaller than China’s, and there are no notable weekly shifts.
What these figures do not prove
- This is one network of one service, not a whole country. A VPN with a different server footprint will see different numbers.
- This is one week. Shifts of 2-3 points over such a period also happen in uncensored countries.
- The app switches protocols after a failure, and in Iran XHTTP goes first. Protocol comparisons inside a country are therefore not clean.
- Only the most frequent failure reasons are known. Failure reasons in the control countries did not make that list, so the error pattern cannot be compared with Germany.
- Control countries have 7-50 times fewer attempts than China and Iran.
What works and what does not: November 2026 summary
Based on one network over one week, with the caveats above.
Works
- Reality and XHTTP in China: 89-91%, on par with Japan and Hong Kong.
- Reality and XHTTP in Iran: 62-71%, on par with Germany and the US on the same network and 20-29 points below Japan and Hong Kong.
- XHTTP in Russia: 74%, better than Reality there.
Works poorly
- AmneziaWG in China and Iran: 32-34%, most failures due to blocked UDP.
- Any protocol on an Iranian network from which no server answers. Disguise does not help when nothing is let out.
Not valid as evidence
- A weekly dip of 2-3 points as proof of a new block: control countries swing the same way.
- The gap between Reality and XHTTP in Iran as proof that one of them is more reliable.
The status of other protocols by country is collected in the overview of what is blocked and what works in China, Iran and Russia.
One protocol really does behave differently in the two countries, but on this network the difference is only partly about censors. Chinese clients connect like Japanese ones, Iranian clients like German and American ones, only with a heavier failure pattern: tunnels go silent more often, and sometimes nothing answers at all. Whether Reality’s minus 6 in Iran was a fluke, the coming weeks will show.