AmneziaWG in Iran and China: why it connects only now and then
AmneziaWG in Iran and China connects far less often than Reality and XHTTP. How the disguise works, where the connection breaks and what UDP has to do with it.
AmneziaWG connects noticeably worse in Iran and China than protocols hidden inside ordinary web traffic, such as VLESS Reality and XHTTP. An estimate based on user reports and open discussions: on difficult networks, roughly one attempt in three gets through. There are no precise public measurements, so the figure remains an estimate. The most plausible explanation is not that censors have learned to recognise AmneziaWG. It is about transport: AmneziaWG runs over UDP, while Reality and XHTTP run over TCP, like ordinary websites. This has not been publicly proven, it is a hypothesis. But it fits the way failures look better than any alternative.
What is AmneziaWG and how is it different from WireGuard
WireGuard was written by Jason Donenfeld and entered the Linux kernel in 2020. It is fast and simple. It was built for encryption, not disguise. WireGuard never tried to hide from a censor.
AmneziaWG is a WireGuard fork from the team behind the open source Amnezia VPN client. The encryption is unchanged. Only the way packets look from the outside is different. KRAKOBIT has already covered in its overview of what is blocked and what works in China, Iran and Russia how WireGuard and OpenVPN fell to filtering and why AmneziaWG became one of the answers.
How a censor recognises plain WireGuard
Plain WireGuard has three tells, all visible from the first packet:
- every packet starts the same way: a message type number followed by three zero bytes;
- handshake packets always have the same size, 148 bytes from the client and 92 bytes in reply;
- the exchange always follows the same pattern: a short greeting, then a data stream.
A filter needs only a simple rule to catch this traffic. It is like spotting someone in a crowd who wears a uniform and a name badge.
What AmneziaWG changes
AmneziaWG erases these tells one by one:
- before the handshake, the client sends several decoy packets of random length, so the first packet is no longer recognisable;
- random bytes are added to handshake packets, so their size changes every time;
- type numbers in the headers are replaced with arbitrary values chosen by the server owner;
- in newer versions, the first packets can be made to resemble other protocols, for example ordinary DNS queries.
After that, WireGuard cannot be found by signature. Against a censor hunting specifically for WireGuard, that is enough. Against a censor that dislikes any unfamiliar UDP traffic going abroad, the disguise changes little: the packet is dropped without being examined.
How much worse AmneziaWG performs: what is known
There are no shared public measurements of protocols in Iran and China. Censors do not publish their rules, and VPN services rarely disclose their statistics. The picture is built from user reports, discussions on circumvention forums and censorship research.
In summary it looks like this. This is an estimate, not a measurement.
| Protocol | Transport | How it hides | China (estimate) | Iran (estimate) |
|---|---|---|---|---|
| WireGuard | UDP | it does not | almost never works | almost never works |
| AmneziaWG | UDP | erases the WireGuard signature | hit and miss, about 1 in 3 | hit and miss, about 1 in 3 |
| VLESS Reality | TCP | looks like a visit to a real website | works on most networks | works, with interruptions |
| XHTTP | TCP | looks like ordinary web requests | works on most networks | works, with interruptions |
One detail stands out. In Iran every protocol struggles, while in China TCP disguise holds up noticeably better. Yet AmneziaWG, judging by user reports, behaves similarly badly in both countries. If that is true, the cause is most likely shared and does not depend on how sophisticated the rest of the filtering is.
How to read any connection success rate
Even where nobody blocks VPNs, some connections fail: a weak mobile network, overloaded café Wi-Fi, a temperamental router, a server on the other side of the world. So a success rate in Iran or China says little on its own. It has to be compared with the same protocol in countries without censorship, such as Japan or Germany.
The rule is simple. If a protocol succeeds in China as often as in Japan, censorship barely touches it. If the gap is tens of percentage points and holds for weeks, it is no longer chance or bad servers.
Where exactly a connection can break
To the user, a failed connection always looks the same: it does not connect. But for the owner of a VPN service it matters at which step it broke. There are three steps, and each means something different.
| What you see | What it means | Whose weak spot |
|---|---|---|
| No handshake, the server never answered the first packet | the packet was dropped on the way and never reached the server | UDP protocols only, AmneziaWG included |
| Connection is up, but nothing flows through the tunnel | blocking by address after the session starts, heavy throttling, a cut after a few seconds | every protocol |
| No server answers over any protocol | a whitelist regime or a shutdown of external internet on this network | every protocol equally |
The first row is the most revealing. If AmneziaWG most often gets no reply to its very first packet, the filter never even tried to figure out what the traffic was. It simply dropped a UDP packet headed for a foreign address. Header disguise cannot help in that case by design: the filter never gets as far as the headers.
According to user reports, this is exactly how AmneziaWG fails most often in China and Iran. That is the basis of the UDP hypothesis.
Why it is easier for a censor to close UDP than TCP
It all comes down to the cost of blocking. TCP on port 443 is the entire web: banks, marketplaces, government services, clouds, work tools. Cutting it crudely means breaking the country’s own economy. So a censor has to hunt on TCP with precision: check which site a connection goes to, compare behaviour with a normal browser, probe suspicious servers.
UDP to foreign servers is a different story. Here is what breaks if it is closed:
| What runs over UDP | What happens if it is closed |
|---|---|
| DNS, the address book of the internet | queries can be sent to local servers, users will not notice |
| The browser’s fast mode for loading sites | the browser switches to the normal mode on its own, pages load a little slower |
| Calls in foreign messengers | in Iran and China those messengers are already restricted or blocked |
| Online games | annoying, but players can be pushed to local servers |
The collateral damage from closing UDP is small. The censor does not need to recognise AmneziaWG. It is enough to drop anything unfamiliar flying abroad over UDP. Against such a filter, disguise is powerless.
There is a caveat. Neither China nor Iran has announced that it cuts UDP. This is a conclusion from indirect signs, not an established fact.
China: what research shows
The Great Firewall can block traffic that looks like a random stream of bytes with no recognisable protocol. GFW Report researchers described the mechanism in a paper at USENIX Security 2023: according to them, such blocking began in November 2021. That work was about TCP connections. Whether the same approach is applied to UDP has not been shown publicly. But AmneziaWG decoy packets with random content look exactly like what such rules are looking for.
A second study by the same group, presented at USENIX Security 2025, showed that since spring 2024 the firewall has been parsing the first packets of QUIC, the fast web protocol that runs over UDP, and reading the site name from them. The takeaway for AmneziaWG is unpleasant: China has invested in inspecting UDP traffic. If you disguise yourself as something familiar, the fake has to be very convincing, or it becomes suspicious in its own right.
Iran: internet in tiers
In Iran, WireGuard and OpenVPN stopped working en masse during the protests in autumn 2022. Since then the authorities have refined tiered access, where some networks only let approved traffic out. In June 2025, during the war with Israel, international traffic was almost completely cut.
On such networks any VPN is useless, and AmneziaWG is no worse than Reality there: nothing answers at all. That is why tools of a different class appear in Iran, ones that need no server of their own, such as Shroudly, which KRAKOBIT covered in early November.
Why Reality and XHTTP get through more often
VLESS Reality runs over TCP and from the outside looks like an ordinary visit to a large, real website. According to the Xray project’s description, if a censor probes such a server itself, it is answered by the very site the server is imitating. Blocking Reality crudely would mean cutting off a noticeable share of the foreign web.
XHTTP splits traffic into ordinary web requests. It can be routed through the same content delivery networks that thousands of regular websites use. Blocking it then hits them too.
That does not make them invulnerable. In Iran both protocols work with interruptions: the connection comes up, then traffic stops flowing. That is what blocking by address and throttling after connection look like. Fighting TCP disguise is more expensive and slower for the censor, but it does fight. More on which protocols currently hold up in each country is in the overview of VPN in China, Iran and Russia.
Why AmneziaWG is still worth having
The protocol has not gone anywhere and does its job. It was built against filters that look for the WireGuard signature. On those networks the disguise works: the signature is gone.
It keeps honest advantages over TCP disguise. It is faster, has lower latency and usually drains a phone battery less. For games and calls UDP is simply better suited: a lost packet does not stall the whole stream.
For the owner of a VPN service, the practical point is this. AmneziaWG makes sense as one of several options, not the only one. Where UDP abroad is allowed, it gives the best speed. Where it is cut, the app needs to switch quickly to a TCP protocol.
What in this picture is hypothesis and what is fact
Fact: AmneziaWG removes the WireGuard signature, and it cannot be found with a simple rule. Fact: China can block traffic that looks like random bytes and inspects part of UDP traffic. Fact: Iran has networks where nothing is allowed out.
Hypothesis: the main reason for AmneziaWG’s low success rate is that UDP to foreign addresses is dropped before the handshake. It explains why the protocol performs about equally badly under two very different censorship systems. Only large measurements across networks and operators can fully test it, and none are publicly available yet.
AmneziaWG won the war against signatures. In Iran and China, by all appearances, a different war is under way, against UDP itself, and in that war header disguise decides almost nothing.