For years, behavioural analysis was the clever answer to bots: even if a script could submit a form, it could not move a mouse like a person or type with human rhythm, so you watched the behaviour and caught the machine. In 2026 that advantage has largely evaporated. Industry research now shows bots that reproduce human-like cursor movement, natural typing cadence and realistic session timing well enough to fool the behavioural tools built to catch them. The tell we relied on has been learned by the other side.
Why behaviour stopped being proof
Behavioural detection rests on an assumption that no longer holds: that human-looking behaviour requires a human. Cheap automation and better models have broken it. A modern bot can wander a cursor, pause "to read", and fill a field at a believable speed. Worse, the rise of legitimate AI agents means human-like automated behaviour is now sometimes exactly what you want to allow — so even a perfect behavioural detector would be solving the wrong problem. Behaviour tells you less and less about whether to trust a session.
The signal that did not degrade
Here is what a decade in this work has taught me: when one signal gets cheap to fake, move to the one that is expensive to fake. A bot can simulate a human's mouse for almost nothing. What it cannot cheaply escape is the need to connect from somewhere. The network behind a request is a physical fact, not a rendered behaviour:
- A real person is on a home or mobile connection, doing one session's worth of work.
- A bot operation runs from hosting and datacenter networks, or hides behind VPNs and residential proxies — and betrays itself in the aggregate: many sessions, one network; impossible geographies; velocity no household produces.
You cannot read that from how the cursor moved. You read it from where the traffic came from.
Layer, don't replace — but lead with the network
Behavioural signals still have a place as one input among several. But they can no longer be the thing you trust most, and the network should be. Lead with a network verdict, corroborate with payment signals like the issuing country, and treat behaviour as a tiebreaker rather than a gatekeeper. Fraudex gives you that leading signal for any IP in a single call — network, type and a bot flag — so a session that behaves perfectly but connects through a proxy still gets caught. See the VPN detection API and the proxy detection API.