Stateful firewalls used as censorship middleboxes exhibit counter-intuitive implementation behaviors: FW-3 forwards ACK packets before a TCP handshake is initiated, and FW-1 actively spoofs RST packets in response to unsolicited traffic to thwart evasion attempts. These vendor-specific quirks create or close evasion opportunities that are invisible to rule-verification tools and not predictable from policy documentation alone.

Evasion attacks generated against one firewall-deployment combination do not transfer well to other settings: a deployment-agnostic approach (used by censorship circumvention tools) fails to generate effective attacks across diverse victim stacks and attacker capabilities. Pryde's deployment-aware, modular workflow finds successful attacks across configurations with and without insider threats, and against multiple attacker success criteria (data delivery vs. victim ACK vs. attacker receipt of ACK).

§1 Introduction (Finding 7), §3.2 evaluation

TCP-compliant packet alphabets are insufficient for modeling stateful firewall evasion. Including non-TCP-compliant traffic — specifically flipped-direction SYNs, out-of-window seq/ack numbers, and packets that form a parallel TCP connection in the reverse direction — is what unlocks discovery of deep attack paths. Prior model-inference work (Alembic) that restricted itself to compliant sequences produced models incapable of generating any of the 6,000+ attacks Pryde found.

§4.1–§4.2 detection

Pryde generates more than 6,000 successful and unique evasion attacks against 4 popular stateful firewalls, which is 2–3 orders of magnitude higher than censorship circumvention algorithms (e.g., Geneva) and black-box fuzzing. The gap arises because circumvention tools only uncover shallow evasion sequences and cannot systematically explore the full attack-state space.

§1 Introduction, Finding 1 evaluation

The server-side variant of the blind VPN inference attack—where an in/on-path adversary exploits predictable NAT assignment and tunnel routing semantics to inject spoofed packets indistinguishable from legitimate encrypted traffic—has remained unacknowledged and unmitigated across all tested platforms since its concurrent disclosure in 2019. Unlike the client-side variant, which received partial fixes from Google (CVE-2019-9461, CVE-2024-49734) and Apple (iOS 17.2.1), no vendor has proposed a viable remediation or claimed ownership of the server-side attack surface.

2026-tolley-architectural §2.1, §3.1, §3.4 traffic-shaperst-injectionmiddlebox-interference

Russian TSPU devices directly block ECH by dropping ClientHello messages that contain both an ECH extension and the outer SNI hostname "cloudflare-ech.com" — the static outer SNI Cloudflare advertises in all its ECH configurations. Blocking affects both TLS and QUIC. ECH connections to servers with Cloudflare ECH support but outside Cloudflare's official IP ranges are NOT blocked. TCP segmentation alone or TLS record fragmentation alone did NOT bypass TSPU ECH blocking, but combining both techniques did circumvent it. TSPU has also added TCP reassembly capabilities that defeat previously effective fragmentation-only bypasses.

2025-niere-encrypted §5, Figure 5 dpirst-injectionmiddlebox-interference

The GFI's HTTP and HTTPS filters are now stateful (requiring initial SYN packet with matching sequence numbers) and have been activated on all TCP ports—not only standard ports 80 and 443 as reported by prior studies. This is a significant departure from previous work that found stateless HTTP/HTTPS blocking limited to standard ports. The HTTP filter injects a 403 Forbidden blockpage (not RST packets as used by the GFW), while HTTPS injects a single RST+ACK packet. The GFI also exhibits TCP non-compliance (not requiring a full three-way handshake to trigger filtering), enabling outside-in measurement without in-country servers.

2025-tai-irblock §2.1, §3.2 dpirst-injectionpacket-injectionmiddlebox-interference

China's Great Firewall showed anomalous inconsistency: 13 test vectors produced mixed outcomes—TCP RST injection on some executions and a clean server response on others—with circumvention rates between 10% and 35% across 100 executions per vector. The authors attribute this to heterogeneous GFW infrastructure components applying different HTTP parsing logic, a departure from the GFW's usual consistency.

2024-m-ller-turning §5.2 dpirst-injectionkeyword-filteringmiddlebox-interference

In Brunei, censorship is confined to AS10094, which serves approximately 70% of the country's Internet users. The censor injects RST packets bearing a distinctive fingerprint — the censored query's IP ID field — in response to HTTP requests containing censored Host headers, and censors on all ports without residual censorship. A SYN followed immediately by a PSH+ACK with a censored payload is sufficient to trigger blocking without a completed TCP handshake.

2023-nourin-detecting §3 rst-injectionpacket-injectionmiddlebox-interference

Censoring middleboxes' TCP non-compliance — specifically, their willingness to censor bidirectionally without completing the three-way handshake — enables external vantage points outside a censoring country to trigger and measure censorship without any local endpoint participation. The approach requires only a confirmed censored domain per AS, evidence of bidirectional censorship, and minimal residual censorship.

2023-nourin-detecting §2 middlebox-interferencerst-injectionmeasurement-platform