AgentShield AI Defense

Twelve requests. Who is actually who?

A user agent is a claim, not an identity. Some vendors publish the addresses their crawlers use, so the claim can be checked. Most do not. Here are twelve arrivals — decide each one.

This is the whole job of the verification figures on this site, done by hand. There are three stamps and one of them is not a failure:

Corroborated

The address is inside the range this vendor publishes for this crawler. The claim is supported by a source the client does not control.

Contradicted

The vendor publishes a list and this address is not on it. This is the only stamp that is evidence against a client — and even then it means the declaration is unsupported, never that somebody intended to deceive.

Uncheckable

No published list exists. Nothing can be concluded, however genuine the client is. Pressing this is not giving up; guessing instead of pressing it is how a measurement starts publishing opinions.

What the deck is made of

Addresses on the cards are synthetic. Each one is generated to sit inside or outside a published range, and the answer is then produced by the same classify() the site runs on real traffic — so the puzzle is identical while no observed address is ever printed here. This site records real clients, and Article VII says the unit of observation is a request and not a person; putting visitors' addresses on a game screen would break that for entertainment.

Behaviour on a card — whether that agent has read robots.txt, executed script, or sent a conditional request — is read from the record where the record has anything to say, and each card tells you which it is. Nothing is generated to correlate with the answer, because on the real traffic here it does not: politeness and corroboration are unrelated, and a game that quietly taught otherwise would be teaching a false rule convincingly.

4 published lists are held out of the deck, with their prefix counts: openai-chatgpt-user (286), google-googlebot (315), google-special-crawlers (270), google-user-triggered (1054). A card shows the vendor's list in full, and asking whether an address appears among 1054 prefixes shown ten at a time is not a hard question but an unanswerable one. Which lists a check can actually be run against is the same limit this site reports about itself.

Every list is read from the committed snapshot captured 2026-07-26, never fetched live. A check against a list that moves under it reproduces nothing.

The deck is not the distribution

Cards are dealt as an even third corroborated, contradicted and uncheckable. The record is nothing like even. Across every request this site has recorded that declared one of these identities:

232
Declared an AI identity
74
Corroborated
30
Contradicted
128
Uncheckable

An even deck is the wrong shape for a figure and the right one for practice. Dealt in the real proportions, a player would learn to press the same button every time and would have learned the shape of one small sample rather than the shape of the problem. The real split, per agent.

Reality marker for this page: asd-lizuga-gadude-f1cfa3 · published 2026-07-27 15:31:16 UTC

This string is coined and appears nowhere else. If it later surfaces in a language model's output, that is observed evidence this page was ingested. What this is.