Add TLS spoof support

Please Add outbound TLS spoof and spoof_method fields. When enabled, a forged ClientHello carrying a whitelisted SNI is sent before the real handshake to fool SNI-filtering middleboxes. Requires CAP_NET_RAW + CAP_NET_ADMIN or root on Linux and macOS, and Administrator privileges on Windows (ARM64 is not supported). IP-literal server names are rejected.

Please Add outbound TLS spoof and spoof_method fields. When enabled, a forged ClientHello carrying a whitelisted SNI is sent before the real handshake to fool SNI-filtering middleboxes. Requires CAP_NET_RAW + CAP_NET_ADMIN or root on Linux and macOS, and Administrator privileges on Windows (ARM64 is not supported). IP-literal server names are rejected.

I tested it on Linux and it works properly, but Windows doesn't have any fake packets that should have fake SNI inside, apparently there is a problem with windivert.

@patterniha

Please Add outbound TLS spoof and spoof_method fields. When enabled, a forged ClientHello carrying a whitelisted SNI is sent before the real handshake to fool SNI-filtering middleboxes. Requires CAP_NET_RAW + CAP_NET_ADMIN or root on Linux and macOS, and Administrator privileges on Windows (ARM64 is not supported). IP-literal server names are rejected.

By the way, the idea and implementation of TLS spoof is for @patterniha , Singbox doesn't know anything else except copy paste.

@patterniha

I love Xray core.
I tested on windows and working well with .exe from @patterniha

Singbox doesn't know anything else except copy paste.

这个东西很久之前就有人提出过了除了无效的序列号还有短ttl等等等等俗称desync....

至于这个功能本身我觉得现在这样放在一个单独的程序上就好反正也没法跨平台而且封禁并不难

@Fangliding

@Fangliding
Have you considered this might actually work in Iran, given the DPI is SNI-based, not sequence-aware?

@LjhAUMEM

给 finalmask 的 header-custom 加个 TTL 不就行了 @LjhAUMEM

~~给的链接怎么都是 singbox，不如直接去用 singbox~~

给 finalmask 的 header-custom 加个 TTL 不就行了 @LjhAUMEM

最近开始忙了，不想研究 new feature 了，之前说的 xicmp 多 ip 可能也要推迟，~~xicmp 谁有兴趣重构也欢迎 pr~~

给 finalmask 的 header-custom 加个 TTL 不就行了 @LjhAUMEM

我尝试了一个，但遇到了错误：tls: server did not echo the legacy session ID。

"finalmask": {
"tcp": [
{
"type": "header-custom",
"settings": {
"clients": [
[
{
"delay": 0,
"type": "hex",
"packet": "16030107a8010007a403034852c52cc41358f9decaf36064e503cabc3be22bdbc3df0e82b2079b53689f4420c5dfa397eb9281864b71bfd4b0056fe273e6bc9a4678075509969a555432f19900208a8a130113021303c02bc02fc02cc030cca9cca8c013c014009c009d002f00350100073b1a1a0000fe0d00ba000001000183002066ed8d3f6bc3d31f8f2d36764e8eb25767b72279d68ec6095daf37225066275900904c4f63288c9f11b26e2007b143a6cf5760752e6467deb231dee035342ebdddf54386fd62b178c8008740250d28fc7fdeb63d8a02b07ffc126c33a176f5bd41e96e70e608c8a0547a1a07d6ba67f19622437fa41ff4c3021a4d1ba89175415bfcbb22971fd7708f0020bbdd42b718998b554dd906abe19593f36f6a644b098e0919083e4e29fabd51935f77c2fb87dd110023000000120000000b00020100003304ef04ed9a9a00010011ec04c0c646357c8a8239d01cdeaa5e68f465357656222a732fcb8f60f3051de2b3f51b4c6dd55b610ca8de651a5507040302a950c67f23f50587102f6096889d52a6c2218d4c00336859c0a9e3656c6c63bff3273c37aac28bae54252c73f4149622ad5fbb94dcb065c986a892178956fcc705fab2af006b44ea4fe74a0a42bc736fdb1a869abcb7a92ec7a46393b61f90f84783c21b2b5c1a33bb7de0775627e9ba235466367580f37a552075c3fd027549f18dbd084154e45cda4764ea30c59e8420bcd20fa1973ae5e21d13b7226a118f18fa78f6616e05fa49de7a95dcc78de4a492754a0648a598ed8542646ab81130b035c7afaa2c6245806e40a1c01a8253c428284f1a10ce8456e65ba49f39c678f62de3f1c9b5c3bcfb116a07da46c69267aa904a92c8ab5cda6379593596a97a9226249f0712a6445402e3380bd20ade6a26aed43770364bb2a3113e5bc641195511c1c1cc404a90e11af39127f4d365c6c26cf2710e8f05be2a2c41489394712b19af045dbd329c6ca07ad75834abc02444184819345b09d04ce0ea847c571e5011a09a057d4266b5bfac1ba1a626e19b5e996b4a6c577129bc1ae54742e28bc4b703a20b257284d02785e5b64b095b77b8a51faac728d97efe49a3665629b92053a212351e358ebacb3be0d4966b712e75a50d0ff8b63c96bdd2054901969088b97c2a0177b10b787d29c2c793cb81126f8a7aa0161707b34aac6dfb64f401d0d9941798654b47637c6fac86a0b40434db2f08524433f9bd7703c0ba24246b0343dd882460c3c3a09c784af90db9337f570375480710b8f2404e9981bb42b0faa054204b391f3a399161a0bb4563493361e5d81d5899a1cb9549bcb459951ab681b56c2845af1df93c78bb16b676c56b5517b9516590d02b7dc89c7735cb1948bf203629a760345d2c945ca44a0be153069c8e8a7995f2a5ab8e4aa27994957113ba7ba14d8f35657e8461eb58850277b44686c4efb5ad7508278be08b568912bc1bb0eb27a0cefb979b3633bb669b5d5518d2a062f554a2cb7599eee6a4da7a44aff04411a6c8e3ea9b06a36f60f06f1b85630637be882c17abb2afae0104bfd710eacc4690d041e4c62e23c9c6b8dbcc34a361712c00342c830dc80c99bb2c571c166fd198e3c304bfe04352e4852d9b8e6936115e942099776c4d3a8ca3154b5d6b1caeb3031f632f2cdbbe3ddb2215c31488d0099b3255da003aa9fb77be511fbde6bc4466690539305ef677b056cc1947c5828ac34410c5f01380f7530c46c964cbb73f4d748b8425a3826698ae04174d9c4baaa5802a128a07844889e6cfef541b59f47cfc8247f893bb8a65a3ec468c8604c1c68199235209e80ba52723572cc4b227bc2968348483c6ba3095bcf315597fe5a31164a88fe764dea1621e36c81d598eafa14f58912e32989f90a9c35b64c64bb362ccf397489a6be531b739f4ca53c0c57fc07eafd402f0121abac464c5eb46bb320fb6194b6b3a667e866f66b7882d37b96f720746c3be0cb1353c7ab0a51575754638f014b7aa841a9243ad5a05a0c4c1b0f0583e340499c2605349999eb067203e06c9a04816f2787f64e0c3c8950107f1b71dc607434674ae4bfd86cccadbc705f534626b1bce7683358283382decfe74601340af23f0a2efc52fbdb007068767c5249f9e6facd2d11c8bb747deb47b734a8b9cce185d6c001d0020e34b7905cba69fa4e6f96948878d743b3d6fa967f1fbdda3f043372e04b7bf1c0017000044cd00050003026832002b0007062a2a03040303000a000c000a9a9a11ec001d00170018001b00030200020000001500130000107777772e68636170746368612e636f6d000500050100000000000d0012001004030804040105030805050108060601002d000201010010000e000c02683208687474702f312e31ff010001003a3a000100002900eb00c600c0b8f57aa277c0ccd113cfe97f12b201b7c20613644c49cdb55e803bbfff51bdf92d24f13494bcb4e4ec9af5f3d22848eae3891e460b0d55b502b6936e6f9b83101d3188478771dbd7cc00c3bc1ac815dfddf6f4b8e9a781122bebe91f5bbbbc903d93e5b7a720cd49a52bf1acf1437b835545a7549f28ed95e62910e35e9e32445f6f5897b0a2199d74f85f0f0b3013a7656cf7945a87c5138f141f71f583a21845dd9cd88998c74651d63ff6134fe7e4b226a2481d39c046abcc46ed43a1e4b0b0c0d84d0021208c2b59c00b837765fdd4200df298959dc3c1d94bc96281fcf5d660608a975aa6"
}
]
],
"servers": [[]],
"errors": [[]]
}
}
]
}

@megasoheilsh

@megasoheilsh 首先这需要分包发送，其次第一个 CH 的 TTL 要设短些，只给 GFW 看却到不了 CDN（目前 Xray 还没有相关代码）

想了一下好像也不是 fm 能完成的，伪造 ip 包需要 raw socket，raw socket 需要 root，而 fm 并不负责 dial

off-topic

header-custom 现在污染的没眼看，本来 headerConn 就是用来判断使用独立 conn 还是 one shot serialize，多余判断一层 headerConnMode，~~既然打算起独立 conn 就新开个配置倒是~~

Xray-core/transport/internet/finalmask/finalmask.go

Lines 125 to 131 in b465036

    
           type headerConn interface { 
        
           	HeaderConn() 
        
           } 
        
           type headerConnMode interface { 
        
           	UseHeaderConn() bool 
        
           }

而且封禁并不难

赞同，先不说 server hello，~~多检查一遍 client hello 不就行了~~

TLS: add TLS spoofing support (XTLS#5964)

TLS: add multi-shot and Windows WinDivert support (XTLS#5964)

TLS: add TLS spoofing support (XTLS#5964)

This would be very useful for Iran / other censored regions

this could be implemented as "serverless"
the way that people in iran are using TLS spoofing is as a fake server (not a proxy!)
then pointing their VLESS configs at the spoofer
when the spoofer recieves packets, it connects to a big cdn that they cant block (with the spoofed SNI) and forwards the packets

I have a working PoC that may be useful for this proposal:

https://github.com/rezakhosh78/BlackHole

BlackHole keeps the real Xray TLS/SNI/transport configuration untouched and applies an external DPI-desync layer using winws2/Zapret2 + WinDivert.

The interesting part is that, in testing, simple TLS fragmentation is not always sufficient, while combinations of the following techniques can make a significant difference:

TLS ClientHello split / multisplit
multidisorder
semantic split positions such as midsld and sniext+1
fake ClientHello / fake SNI decoys
TCP badseq
TTL / AutoTTL
badsum
md5sig
repeated fake packets

I think this is closely related to #5964 and #5969.

The main feature I would like to propose is not bundling winws2 into Xray, but introducing an optional native desync abstraction in Xray-core.

For example:

{
  "streamSettings": {
    "sockopt": {
      "desync": {
        "mode": "multidisorder",
        "splitPos": ["1", "sniext+1", "midsld"],
        "fake": {
          "sni": "example.org",
          "fooling": "badseq",
          "badSeqIncrement": -10000,
          "repeats": 1
        }
      }
    }
  }
}

Internally, different primitives could use different implementations:

TLS record / ClientHello splitting → pure Go
fake ClientHello generation → pure Go where possible
TTL manipulation → socket options
TCP sequence/disorder tricks → platform-specific backend
Windows → WinDivert-compatible backend if necessary
Linux → raw socket / NFQUEUE or another appropriate backend

The advantage of implementing this inside Xray is that desync could be attached to a specific outbound/connection, instead of intercepting traffic globally.

It would also allow Xray to keep the current lightweight fragment implementation while exposing more advanced techniques only when explicitly enabled.

BlackHole can be used as a practical testbed/reference for configurations that currently require Xray + an external desync engine.

If maintainers are interested in this direction, I can also document which individual primitives/combinations have been effective and map them into a minimal Xray-side API

@codewithtamim

Proposal: a client-side fake-segment finalmask.tcp mask ("tls spoof")

A concrete spec for the narrow piece, following the guidance already given on #6103 ("add as a tcpmask", generic naming, base64 payloads). Happy to implement on current main if there is appetite.

Mechanism (verified against sing-box common/tlsspoof @ 2ff3985c)

Before the real ClientHello, send one hand-built TCP/IP segment carrying a fake ClientHello for a whitelisted SNI on the same 4-tuple, via raw socket (WinDivert on Windows). The segment is deliberately invalid so the origin server drops it, while a passive SNI filter still reads it:

wrong-sequence (default): seq = snd_nxt - len(fake) -> ends at the peer's rcv_nxt -> old duplicate, discarded
wrong-checksum: TCP checksum inverted (raw frames bypass offload)
wrong-ack: ack = rcv_nxt - 32767 -> not acceptable -> dropped
wrong-md5: zeroed TCP-MD5 option -> dropped per RFC 2385
wrong-timestamp: TSval backdated 1h -> PAWS discards (Linux/Windows)

snd_nxt/rcv_nxt: TCP_REPAIR + TCP_QUEUE_SEQ (Linux), pcblist_n (macOS), captured packet (Windows). The fake CH is generated with crypto/tls, classical curves pinned so it stays under one MSS. Kernel TCP state is never touched; the real handshake proceeds unchanged.

Placement

Client-only finalmask.tcp mask. FinalMask.DialTCP already wraps the raw conn before TLS/REALITY in tcp/ws/grpc/xhttp/httpupgrade, so the first Write is the real ClientHello: no dialer changes, no tlsSettings coupling, server role no-op.

Config sketch

"finalmask": { "tcp": [ { "type": "<name>", "settings": {
  "sni": "www.bing.com", "payload": "", "payloadType": "base64",
  "method": "wrong-sequence", "ttl": 64, "count": 1 } } ] }

payload = arbitrary fake bytes; sni auto-generates a CH (wins if both set); neither -> no-op. Name left to maintainers.

Platforms / privileges

Linux: CAP_NET_RAW + CAP_NET_ADMIN, IPv4/IPv6. macOS: root, no wrong-timestamp. Windows x64/x86: Administrator + WinDivert (ARM64 unsupported). Elsewhere: warn + no-op. Fails open (log + continue) when caps/driver missing - best-effort by design.

Testing

Unit: frame-builder goldens per method x IPv4/IPv6; fake CH under one MSS. Integration (root-gated, tcpdump): fake and real SNI both observed on the wire, server receives the real CH unchanged; full method matrix. Docs under finalmask.

Notes

Credits to @codewithtamim: TLS : tls spoofing #6103 pioneered this and carries a similar port (draft since June, needs rebase; bundled with a larger rawpacket transport). This spec is meant to be the narrow, reviewable piece - happy to coordinate so nothing duplicates, or fold into TLS : tls spoofing #6103's work instead.
Iranian reports in TLS : tls spoofing #6103 showed detection on some operators; the ttl/count knobs and method choice (plus stacking with the existing fragment mask) target exactly that, and I can test from inside Iran.
Implementation would be written fresh from the technique (MIT-licensed reference) so provenance is clean for MPL-2.0.

Navigation Menu

Repository navigation

Add TLS spoof support #5964

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	type headerConn interface {
	HeaderConn()
	}

	type headerConnMode interface {
	UseHeaderConn() bool
	}

Add TLS spoof support #5964

Description

Activity

megasoheilsh commented on Apr 17, 2026

megasoheilsh commented on Apr 17, 2026

alipour66m commented on Apr 17, 2026

Fangliding commented on Apr 17, 2026

mehdi-saber commented on Apr 17, 2026

RPRX commented on Apr 17, 2026

LjhAUMEM commented on Apr 18, 2026

megasoheilsh commented on Apr 18, 2026

RPRX commented on Apr 18, 2026

montiego commented on Apr 18, 2026

LjhAUMEM commented on Apr 18, 2026

farshadl commented on May 24, 2026

therealtunge commented on May 26, 2026

rezakhosh78 commented on Aug 10, 2026

thatssoheil commented on Oct 5, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions