Residential proxy pools are pushing edge defence down to the TCP handshake

The signal exists because the standard left room for it. RFC 793 and the RFCs after it pin down TCP's semantics and hand option sequencing, default buffer allocations and window scaling coefficients to whoever writes the kernel ^[12]. A SYN frame therefore carries the implementer's choices, and the fields an edge can read off it include TCP option ordering, initial window size, TTL hop distance and MSS constraints ^[11]. Reading them during the initial SYN exchange is passive OS fingerprinting, p0f style ^[9]. A headless Chromium can forge its user agent, its screen dimensions, its WebGL vendor string and its TLS client hello ^[7], but the kernel still writes the handshake underneath.

Cardinality is the reason to prefer the kernel field over the address. The pool in the post's diagram is 100,000-plus addresses in front of one identical kernel engine, given as Linux 5.15 headless ^[4]. One fingerprint therefore spans the whole pool, turning 10^5 tracked keys into one ^[1]. Renting those addresses costs the operator orders of magnitude less than the stateful tracking a defender stands up against them, token bucket or leaky bucket or sliding window counter alike ^[6].

The diagram only works if the SYN the edge inspects comes from the machine running the headless browser. A proxy that terminates TCP and opens its own connection to the target puts its own stack on the wire, and the fingerprint then describes the residential device, not the scraper's container. Forwarding at L3 preserves the signal; terminating TCP destroys it. That distinction decides whether any of this transfers, and it is the proxy product you are up against that sets it.

The second limit is resolution. The key identifies a kernel ^[2]. A legitimate visitor whose stack emits the same defaults lands in the same bucket as the fleet, which is why the decision rests on the L7-to-L4 mismatch and not on the fingerprint on its own ^[8]. For that check to pay on your traffic, claimed OS and observed kernel have to disagree far more often in abuse than in ordinary requests, and the base rate depends on who your users are. The post presents the mismatch as an annotated illustration and does not include measured detection or false-positive rates, or a named deployment ^[13].

Evasion is cheap for some adversaries and awkward for others. The post lists three routes to forging the network-layer parameters: raw socket privileges via CAP_NET_RAW, custom eBPF packet generation engines, or non-trivial guest kernel modifications ^[10]. Each needs control of the kernel that emits the packet. An adversary already running a cluster orchestrator across their own container hosts controls the kernels on those hosts ^[3].