How the Great Firewall of China Works in 2026 — A Technical Explainer
By Mehmood Ali · · 14 min read
How China's Great Firewall works in 2026: asymmetric blocking, QUIC SNI censorship, the Wallbleed memory leak and the leaked Geedge Networks files.
The Great Firewall of China (GFW) is the largest network-censorship system in operation, and it has changed a great deal in the last three years. This post goes through what peer-reviewed research had established by April 2026. That covers asymmetric filtering, the first nationwide deployment of QUIC censorship, and a buffer over-read in the DNS injectors that leaked their memory. It ends with the documents leaked from Geedge Networks and the MESA Lab, the two organizations that build the system.
The GFW is not one firewall — it is thousands of middleboxes
The name suggests a single appliance on a border router, but the GFW is a distributed network of middleboxes installed at China's international gateway routers (IGRs) and at the boundaries between provincial networks and ISPs. Most of these middleboxes sit on-path rather than in-path. They tap a copy of the traffic and inject forged responses, but they don't drop packets inline. That matters to circumvention researchers, because a legitimate response that arrives first can sometimes beat the forged one.
The S&P 2025 paper "A Wall Behind A Wall" (Wu & Zohaib) measured the GFW as asymmetric, where earlier work had generally assumed it applied the same rules in both directions. Wu and Zohaib tested 10,000 domains and found 68 that were blocked only when probed from inside China, among them google.com, nyt.com and docker.com. Other domains were blocked only from outside, which suggests separate inbound and outbound blocklists. The paper also documents a second filtering layer in Henan province that inspects only outbound provincial traffic, the first regional firewall known to operate in China.
QUIC censorship — new as of April 2024
Until 2024, QUIC (the UDP transport under HTTP/3) was largely left alone. That changed on April 7, 2024, when the GFW started decrypting QUIC Initial packets at scale and checking their SNI against a blocklist. Once it sees a forbidden SNI, it drops every later UDP packet with the same server IP, destination IP and destination port, and that residual blocking lasts more than 100 seconds.
The implementation has documented weaknesses. It tracks flows by UDP 4-tuple with a 60-second timeout instead of following Connection IDs. It doesn't reassemble QUIC Initials that are split across several datagrams, a gap Chrome's September 2024 changes exploited by accident when they made Initials too large for one datagram. It exempts connections whose source port does not exceed the destination port. Since December 2024 it has also let through QUIC connections that carry an ECH extension. The QUIC blocklist is separate from the DNS one and about 60% of its size by domain count, which points to policy being maintained independently for each.
Zohaib et al. (USENIX Security 2025) give the full measurement methodology and the evidence for 4-tuple tracking. The paper is open access.
Wallbleed — the GFW's memory disclosure bug
At NDSS 2025, researchers from GFW Report, CU Boulder and UMass Amherst published Wallbleed, a buffer over-read in the GFW's DNS injection middleboxes. A DNS query with an oversized label-length prefix made the injector read up to 125 bytes past the end of the packet from its own process memory and copy them into the forged response it sent back. The name is a nod to Heartbleed, and the bug works the same way, only here it sat inside state censorship equipment.
The bug existed from at least October 2021. It was partially patched in September 2022 and again in June 2023, and fully fixed by March 2024. Over two years the team collected 3.1 billion Wallbleed responses. The leaked memory showed Linux on x86_64 (70,497 pointer examples) along with fragments of SSDP, HTTP, TLS, SMTP and DNS traffic, so the middleboxes clearly handle far more than DNS.
One more detail came out of the data. The pool of IPs used in forged responses is fixed and handed out in a set order that repeats. For one query name (4.tt) the researchers recovered 592 addresses used in rotation.
Who builds it? Geedge Networks and the MESA Lab
The September 2025 leak of Geedge Networks and MESA Lab material named the two organizations that develop and run the GFW. Geedge Networks was founded in Hainan in 2018, with Fang Binxing, often called the "father of the GFW", as chief scientist. It sells four products:
- Tiangou Secure Gateway (TSG), the flagship DPI and filtering product. It blocks VPNs and can also inject malicious code into websites and launch DDoS attacks.
- Cyber Narrator, an operator interface for monitoring users in chosen geographic regions.
- TSG Galaxy, the long-term storage and analytics tier.
- Network Zodiac, which monitors infrastructure health and deployments.
MESA Lab sits within the Institute of Information Engineering at the Chinese Academy of Sciences and has run the SAPP platform at national scale since 2014. Geedge products have turned up in Kazakhstan, Pakistan, Myanmar, Ethiopia and elsewhere, so the GFW is now also something China sells abroad.
What this means for VPN detection
Much of what the GFW does to block traffic has a direct counterpart in how commercial services detect VPNs. The clearest case is active probing, which the GFW pioneered against suspected obfuscated servers: when an endpoint handshakes oddly, the censor opens its own follow-up connections and watches whether they are silently dropped. IPLogs and similar services use the same idea from outside China to identify OpenVPN, WireGuard, IKEv2 and REALITY endpoints.
TLS inspection carries over too. Reading the SNI from a ClientHello, or hashing the whole ClientHello with JA3/JA4, is cheap and tells you a lot. The GFW uses it to block, and defenders use it for fraud prevention. Protocol fingerprints work the same way: OpenVPN's HARD_RESET opcode, WireGuard's handshake initiation and IKEv2's SA_INIT each look distinctive on the wire, so a detection engine can recognize them without decrypting anything.
Try it yourself
To see whether a particular IP is a VPN, a Tor exit or a datacenter address, paste it into the home-page checker. It runs our 7-layer detection pipeline and lists every signal that matched next to the verdict.
References
- Wu & Zohaib et al., "A Wall Behind A Wall: Emerging Regional Censorship in China", IEEE S&P 2025.
- Zohaib et al., "Shining Light on the Inner Workings of QUIC Censorship in China", USENIX Security 2025.
- Fan et al., "Wallbleed: A Memory Disclosure Vulnerability in the Great Firewall of China", NDSS 2025.
- Hoang et al., "How Great is the Great Firewall? Measuring China's DNS Censorship", USENIX Security 2021.
- Geedge Networks / MESA Lab leak analyses, September 2025 (see Wired, Amnesty Tech, GFW Report).