9 August 2026
What is active probing?
Most explanations of censorship stop at inspection. The network watches traffic go past and blocks what it recognises. That is the passive half of the problem, and a decent obfuscated transport deals with it.
The active half gets discussed far less and is considerably harder to design around. Having seen a connection it finds suspicious but cannot pin down, the censor goes and connects to that server itself. Then it judges the server by how it replies.
Why this beats a lot of obfuscation
Picture a transport built to look like nothing in particular. Encrypted bytes, no recognisable header, lengths that vary. Against a passive filter this is fine. There is no signature to match on.
Against a prober it can come apart quickly, because the censor is no longer limited to what you sent. It opens its own connection and sends whatever it fancies: random bytes, a replay of traffic captured earlier, half a handshake. Then it watches.
Real web servers have manners. Send one nonsense and you get an HTTP error, or a TLS alert, or a timeout with a particular shape to it. A proxy impersonating nothing in particular usually behaves differently. It might sit waiting for more data, or hang up instantly, or return a length that betrays it was hoping for a valid handshake and did not get one. None of that is a confession. It only has to be unlike a normal web server often enough to justify a block.
There is the asymmetry that makes probing so effective. Passive inspection asks whether traffic looks like a VPN. Probing asks whether a server behaves like the thing it claims to be. Passing the first tells you nothing about the second.
The research on this is public
Researchers have been studying the Great Firewall's probing infrastructure for over a decade, and the papers are worth more than my summary of them.
Examining How the Great Firewall Discovers Hidden Circumvention Servers (Ensafi, Fifield, Winter, Feamster, Weaver and Paxson, IMC 2015) worked from years of server logs and traffic captures. Two findings stand out. The sensors that notice suspicious traffic turned out to be separate from the Great Firewall's main blocking infrastructure, so the watching and the blocking are not the same system. And fingerprinting the probes, down to TCP initial sequence numbers and timestamps, showed shared state across probe sources that otherwise looked unrelated: a large number of addresses driven by a small number of processes.
The paper also measured what probing was worth to the censor. Vanilla Tor was almost entirely blocked for their clients inside China, while obfs2 and obfs3 stayed mostly reachable.
How China Detects and Blocks Shadowsocks (Alice, Bob, Carol, Beznazwy and Houmansadr, IMC 2020) is the sharper picture, and the one I would read first. The Firewall picks its suspects using the length and entropy of the first data packet in a connection, then sends seven distinct types of probe. Some are partial replays of the user's own earlier traffic. Others are random payloads of varying length, aimed at how different Shadowsocks implementations react to input they cannot decrypt. The probers used thousands of IP addresses and again appeared to be centrally coordinated.
Note what the trigger is in that second one. Not a signature, but the size and randomness of your very first packet. Encrypting harder makes the entropy worse, not better.
The GFW Report project has kept publishing measurements since. The consistent finding across that literature is that probing is automated, systematic, and follows the triggering connection closely in time.
What a transport has to do about it
If the censor is going to knock, the server needs to answer like whatever it says it is.
The strongest approach currently is to actually be that thing for anyone who cannot prove otherwise. The server completes a genuine TLS handshake using a real site's certificate chain, and any connection failing to present the right client credentials gets forwarded to that real site. A prober therefore does not receive a suspicious response, nor a clever imitation of a website. It receives a website, because its connection was handed to one. There is no behavioural difference left to detect.
REALITY works along these lines, and it is why a probe-resistant setup can survive somewhere a plain encrypted proxy would be found and killed inside a day.
The other school of thought withholds everything until authenticated: stay silent unless the client proves knowledge of a shared secret in its opening message, so a prober with no credentials gets nothing to fingerprint. It works, though at scale "connections that go nowhere" can become its own signal.
Where probe resistance runs out
Probe resistance is a claim about the probes being sent this year. Techniques change. A design that answers today's probes convincingly may not answer next year's, and treating it as a permanent property is how people get caught out.
It also does nothing outside the connection itself. If reaching a provider's website to download the client is blocked, how well that client resists probing is irrelevant, because you never got it. That is the unglamorous reason "install before you travel" keeps appearing in advice about heavily filtered countries. It is not filler.
And none of it touches volume or timing at the edge. A censor who cannot work out what you are doing can still see that you moved a great deal of encrypted data to one endpoint for six hours.
Encryption strength is the wrong question
If you are choosing a VPN for somewhere that filters seriously, "military-grade encryption" tells you nothing about either half of this. Encryption strength is not the thing under test.
Better questions are narrower. Does the transport survive an active probe, and by what mechanism? Is there more than one, so that a failure is survivable? Does the provider describe limits or only wins? Anyone who has thought about probing can answer the first two specifically. Anyone who has not will start talking about encryption again.
Pangea's cascade includes REALITY for exactly this reason, alongside three other transports it falls back through automatically. The VPN for China page has the specifics, and there are five days free with no card if you would rather test it than take our word for it.