By Evan Vega
Nine new web addresses were each requested by outside software between 1.2 and 4.4 seconds after their security
certificates were recorded in public logs, according to a pre-print published on 13 November 2026 by Guerin Lee Green of
Novel Cognition (doi:10.5281/zenodo.24203364). Nine comparison addresses on the same three domains,
covered only by a shared “wildcard” certificate that names no address, drew almost nothing: eight received no outside
request in 24 hours, and the ninth was reached after 23.9 hours.
Every certificate that browsers trust is published in the Certificate Transparency logs, a public record set up to
catch certificates issued in error. The record doubles as a live list of new hostnames. The usual advice for keeping a
hostname off that list is a wildcard certificate, and the paper says that advice had not been tested with the served
certificate held constant. In this experiment all 18 addresses served the same wildcard certificate and were empty. The
only difference between the two groups was whether a certificate naming the address had been logged.
All 9 logged addresses were requested by third parties within 24 hours, against 1 of the 9 unlogged ones (95%
confidence intervals 70.1 to 100.0% and 2.0 to 43.5%; Fisher’s exact test, p = 0.00041). Within a minute of issuance each
logged address had been requested by at least five separate clients on at least five networks, and within an hour by a
median of 31 clients. On one address a request arrived two seconds before the certificate software reported the
certificate issued. The paper’s explanation is that log readers act on the precertificate, which is logged before the
certificate exists.
Cloudflare’s DNS analytics, which are sampled estimates, point the same way: about 7,280 lookups for the nine logged
names during the window against about 120 for the nine unlogged names. “An operator who plans to secure a host after the
certificate arrives is already late,” the paper says (the pre-print).
A Novel Cognition pre-print put 18 empty hostnames on three domains: nine named in their own certificate and so published in the Certificate Transparency logs, nine covered only by a wildcard certific
The control behaved as the standard advice predicts, with one exception. Eight of the nine unlogged addresses received
no request from anyone but the operator and the hosting provider. The ninth was hit 23.9 hours after its DNS record was
created: one client on a single cloud network sent 499 requests in about eight seconds under 33 different software names.
A second unlogged address on the same domain was hit the same way 130 seconds after the 24-hour window closed, which the
paper reports separately from its main result.
No certificate names either address, so the logs did not disclose them. The paper leaves two candidates: guessing, or
a leak through the author’s own monitoring checks, which had requested five of the nine unlogged names, both of the reached ones
among them. It does not claim to separate the two. It does show that guessing was already under way. The three domains
were answering lookups for non-existent subdomains at a median of 315 to 490 a day each in the eight days before the
experiment, across 459 distinct names, and one of the six labels used in the test had been tried on one domain three days
before the address existed. All six labels were ordinary English words. The paper’s conclusion: a wildcard certificate
keeps a label out of the logs but does not keep a dictionary word from being guessed.
The scanners that follow the certificate logs did not browse. An estimated 92.0% of their requests in the first hour
were for paths that no page listed, and 70.0% of those were for environment, credential and configuration files such as
/.env, which often hold passwords and API keys. With the busiest scanner left out the unlisted share was 64.5%. The
honeypot answered every such request with “not found” and offered no bait.
Volume was concentrated. The paper estimates 21,889 third-party requests from 92 networks in the window; one network
sent 41.0% of them, nearly all from a single scanner, and five networks sent 85.8%. More than half of the traffic to the
logged addresses, 57.1%, arrived in the first hour.
Each address listed 500 URLs in one of three places: a sitemap file at the conventional /sitemap.xml path with an
empty home page, plain HTML links on the home page, or sitemap XML served as the home page. Anthropic’s ClaudeBot reached
all nine logged addresses 161 to 283 seconds after their certificates were issued, asked for robots.txt first and fetched
all 500 URLs on each. That includes the three addresses whose list sat at /sitemap.xml with nothing pointing to it, which
the crawler requested on its own about an hour after issuance.
OpenAI’s GPTBot arrived 1,361 to 5,498 seconds after issuance and fetched all 500 URLs only on the three addresses
whose home page listed them as HTML links. It fetched none from either sitemap layout. OpenAI’s OAI-SearchBot made one
robots.txt request per address. None of the three reached an unlogged address. The pages were empty and the hosts set no
crawler rules, so the paper describes a first pass over content-free pages and does not test whether robots.txt is
obeyed.
Crawlers were counted only when the request came from the operator’s own network, for a reason the data supplied. The
client that hit the unlogged address presented 14 different AI-crawler names, along with Google and Bing crawler names,
from one cloud network in eight seconds. Counting by the name in the request would have recorded 14 AI crawlers on an
address where the network check finds none (the dataset).
The result rests on one 24-hour run: 18 hostnames on three domains, one certificate authority (Let’s Encrypt) and one
content-delivery network (Cloudflare). It does not test hostnames with random labels, which the paper names as the next
comparison, and it does not show that an unlogged hostname stays unfound for longer than a day. The three domains are
written as D1 to D3 in the paper and the data. The request-level dataset, the honeypot’s source code, the analysis plan
and a script that recomputes every number are free to download under a CC BY 4.0 license
(doi:10.5281/zenodo.24203362); the pre-print is at
doi:10.5281/zenodo.24203364.
The paper’s own disclosures: Green operates the three domains, founded Novel Cognition and runs the
Hidden State Drift Mastermind, a paid practitioner group. He is a paying
customer or ordinary user of Cloudflare, Let’s Encrypt, Anthropic, OpenAI, Google and Microsoft services and reports no
other relationship with them. A large language model, Anthropic’s Claude, wrote the honeypot and the analysis scripts and
drafted the manuscript under his direction; Anthropic also operates ClaudeBot, one of the crawlers measured. Novel
Cognition tests research and vendor claims for companies as part-time AI leadership
(novcog.com/contact).
Part of the Frontier Watch Series: Read the previous investigation
More Coverage:
→ Read this investigation on North Denver Tribune
→ Coverage from Daily Colorado News