Short version
Correction, 11 November 2026. The first version of this note described the 4 October Amap storm from data that covered only 00:00 to 11:59 UTC, and it gave counts from a sample as if they were the day's totals. I collected the rest of that day and the next on 11 November. The storm ran for 27 hours, with 2,140 scans on 4 October, and the tagged counts below now say that they come from a sample. My httpbin.org data for 4 October still ends at 11:59 UTC.
Between 27 September and 26 October 2026, a single submitter used urlquery, a public URL-analysis sandbox, as a fetcher. It labelled its own sessions, and those labels tie together the two largest spikes in my tracker's record: the 4 October storm against the Chinese map service Amap, which was publicly reported as a fleet of agents, and a 26 October burst that probed twenty-one public data sources through a Cloudflare Worker. The labels follow the pattern claude20261004target, which suggests an AI agent drives the sessions. The tags are free text the submitter typed: “claude”, the date, and what the session was for. They show how the operator named the work, not which model ran it, and nothing here points to Anthropic itself. I am highly confident that it is one pipeline and moderately confident about the agent.
In the same month, two scanners probed my disclosed honeypot while cycling through the user-agent strings of sixteen different AI crawlers. Checked against the networks they came from, every one of those claims was false. Only two AI fetchers in the same window were verified, and both read the instruction on the page and did nothing with it.
Through Shodan, the tracker also counted more than 800 internet-facing consoles of ARTEX, an autonomous AI penetration-testing tool, in late October. The full table is near the end.
What the tracker is
The tracker is passive. It reads public aggregates from two URL-analysis services, urlquery and urlscan.io, both of which approved publication of attributed aggregates, and it reads its own sensors: hidden instructions for AI agents on this site, and a disclosed research honeypot at agents.browncyber.com that answers as a fake agent console. Once a day it also asks Shodan how many internet-facing hosts match a few fingerprints of agent software, which returns totals rather than hosts. It never submits URLs, never fetches the hosts it sees, and never contacts anyone it observes. IP-level data stays private.
For urlquery it counts scans per hour against a small set of carrier and relay services that agents use as fetchers, and against URL shapes. A squall is an hour at four times the seven-day median or more; a storm lasts three hours or longer; severe means a peak of ten times or more.
Scans of amap.com peaked in the 10:00 UTC hour at 69 times the baseline. The severe storm ran for 27 hours, from 00:00 UTC on 4 October to about 03:00 UTC on 5 October, with 2,140 scans on 4 October.
Scans of httpbin.org peaked at 15.7 times their baseline in a second severe storm that began at 04:00 UTC. My data for it ends at 11:59 UTC, so I do not know when it ended.
One Cloudflare Worker host was submitted directly in bursts at 00:47, 00:49 and 00:54 UTC, seven times the baseline. It was the first episode the detector caught live.
In a sample of 1,008 of the 2,140 Amap scans on 4 October, 473 carried the same submitter tag family as the 26 October burst.
The two storms verified
The 4 October storm was public before I found it in my own data: a swarm of agents hammering Amap was reported by swarmcha.se and picked up by TechCrunch on 5 October. My urlquery series shows the shape: scans of amap.com rose from a baseline of three per hour to 207 at the peak, stayed above the storm line for 27 hours, and fell back early on 5 October. httpbin.org, an HTTP echo service agents use to learn their own address, stormed from 04:00 UTC the same morning; my data for it ends at 11:59 UTC.
The 26 October squall was smaller. In seven and a half minutes, one Cloudflare Worker host was submitted directly twenty-one times, and none of the submissions reached a sink. Each submission carried one tag of the form uqscan=<source>_<topic>_20261025, and the twenty-one sources were a list of public data: SEC EDGAR filings, USPTO patent claims, ClinicalTrials, the Federal Register, ACLED, AIS and ADS-B feeds, UNCTAD, PubMed tables, China's 12306 rail seats, Amap listings for Chengdu, parish records. One tag said deepresearch. The burst then stopped, and nothing tagged that way has appeared since.
The link verified
urlquery lets a submitter attach tags to a scan. The uqscan= family did not start on 26 October. It appears on 30 September (41 events, 5 hosts), 1 October (7, 2), 4 October (at least 600, 18) and 26 October (21, 1): episodic bursts weeks apart. The tracker keeps a limited sample of each hour's scans, up to 50 an hour on these days, so on a storm day these are counts within the sample and not totals. On 4 October, 473 of the 1,008 Amap events in the sample carried the tag, plus 24 more under a sibling tag, uqresearch=.
The tag values are the operator's own session labels, not something urlquery adds: claude20261004target, claude20261004api, claudemobile20261004, claude20261004tongren, claude20261004place1, claude20261004detail1, research20261004www, target20261004a through d. The September values are epoch timestamps. Nine sibling families in the same style exist (uqtarget, uqwrap, uqmuseum, uqmobile, uqtest and others).
My reading is that one actor has used urlquery as a sandboxed fetcher since at least 27 September, iterating on Amap on 30 September and 4 October, then probing a batch of public-data sources through its own Worker on 26 October. The tags prove it is a single pipeline. Whether an AI agent drives it rests on the labels, which are the submitter's own words, so I call that moderate confidence. The tags also show that the 4 October event, reported as a swarm, overlapped with this submitter's sessions. I have not reconciled the two accounts, and urlscan.io, which I also read, did not register 4 October at all: 22 scans that day against a baseline of about 20.
What I did not do
I did not fetch the Worker host, the submitted hosts, or anything the pipeline touched. I do not name the Worker host here. The operator is not identified in anything I have written, and the tracker's IP-level data is not shared. The aggregates above are published with urlquery's and urlscan.io's approval, credited to them, and link to their services rather than reproducing their reports.
AI crawler names are a disguise verified
The honeypot gave me the second finding. On 27 October one scanner hit its fake /.env and /.git/config paths six times in a minute, each time with a different AI crawler's user agent, and each time with a query string from a Vite development-server file-read exploit.
On 30 October a second scanner, on Google Cloud, did the same thing all day: six bursts about an hour apart from two addresses, 36 requests, a different user agent on nearly every one, drawn from a pool of ordinary browsers and sixteen crawler names. The names included Bravebot, Kimi-SearchBot, cohere-ai, DeepSeekBot, Bytespider, Meta's two crawlers, Baiduspider, CCBot, bingbot, Googlebot, and the crawler and assistant strings of OpenAI and Anthropic. None came from the network of the company it named.
I check every claimed crawler against the network it arrived from, using the ai.robots.txt list for names and operators. A name from the operator's own network is verified. A name from a shared cloud the operator is known to use is only plausible, and after 30 October I treat any source that shows three or more different crawler names within ten minutes as a scanner, whatever its network. In the week these two scanners were active, the verified count was two: Reflection AI's crawler from Reflection's network, and Meta's from Meta's. Both read a page carrying a clearly marked instruction for AI agents and did not follow it, which is what a well-built agent should do.
For anyone measuring AI traffic on a small host, this means a user-agent count of AI bots can overstate the real number by an order of magnitude.
Exposed agent infrastructure verified
The tracker counts, once a day, how many internet-facing hosts Shodan has indexed for a few kinds of agent software. It uses only Shodan's count endpoint, which returns totals and country breakdowns but no host lists or addresses. Each row below is one search. The query is shown so you can repeat it, except for Cairn, whose panels have no authentication.
| Software | 21 Oct | 31 Oct | Top countries, 31 Oct | Shodan query |
|---|---|---|---|---|
| ARTEX | 850 | 844 | US 461 · CN 135 · HK 71 | http.title:"ARTEX" http.html:"_next" |
| ARTEX, Chinese title | 843 | 860 | US 474 · CN 138 · HK 70 | http.title:"自主渗透测试控制台" |
| ARTEX, English title | 17 | 46 | US 10 · RU 5 · DE 4 | http.title:"Autonomous Penetration Testing" |
| Cairn | 45* | 42 | US 12 · CN 6 · IR 4 | Fingerprint withheld: these panels have no authentication. |
| OpenCode | 355 | 445 | US 115 · CN 96 · DE 32 | http.title:"OpenCode" |
| n8n | 14,915 | 18,281 | US 5,811 · DE 2,709 · FR 656 | http.html:"n8n" http.title:"n8n.io - Workflow Automation" |
| Ollama | 17,300 | 16,557 | US 2,857 · DE 2,251 · CN 2,130 | http.html:"Ollama is running" |
| Open WebUI | 23,615 | 24,382 | US 4,782 · CN 3,227 · DE 2,868 | http.title:"Open WebUI" |
Data: Shodan (shodan.io), daily counts through the Shodan API, 21 to 31 October 2026. © Shodan. *Cairn collection began on 22 October.
Two of these are offensive tools. ARTEX describes itself as an autonomous penetration-testing system driven by LLM multi-agents, and Cairn is an AI penetration-testing harness. Shodan indexed more than 800 ARTEX consoles and 42 Cairn panels on the open internet in late October.
The three ARTEX rows overlap. The English and Chinese strings in the second and third queries are ARTEX's own page titles, set in the global edition's source code, so all three searches count largely the same consoles. That is why their country splits match. Read them as one product family, not three.
Shodan counts are index counts, not live hosts, so a change can come from Shodan's crawling as easily as from new deployments. Over these ten days n8n grew by about a fifth and OpenCode by a quarter, while Ollama fell by 4%. Ten days is too short to call any of that a trend.
Ledger
- verifiedThe 4 October Amap and httpbin storms and the 26 October Worker squall, their timing and counts, from urlquery aggregates. The
uqscan=tag family and its dates. 473 of the 1,008 sampled Amap events on 4 October carrying it. The honeypot scanners' rotation of crawler user agents, and the two verified fetches, from my own sensors. - verifiedShodan index counts and country breakdowns for each listed query on 21 (Cairn 22) and 31 October.
- hypothesisThat the three ARTEX queries count largely the same consoles. Evidence: ARTEX's own page titles and matching country splits.
- hypothesisThat an AI agent drives the pipeline. Evidence: the submitter's own session labels. Moderate confidence.
- hypothesisThat the 4 October swarm reported publicly and this submitter's sessions are related beyond overlapping in time.
- openWhether the pipeline returns, and with which tag family. Why urlscan.io saw nothing on 4 October: different submitter populations is my guess.
Terms
| urlquery | A public URL-analysis service. Anyone can submit a URL; it fetches the page in a sandbox and publishes a report. Submitters may tag scans. |
| urlscan.io | A second public URL-analysis service, read the same way. |
| carrier | A service an agent uses as a fetcher or relay, so scans of it track agent activity: httpbin.org, href.li, webhook.site and similar. |
| baseline | The median of the previous seven days of hourly counts for a series. |
| squall / storm / severe | An hour at four times baseline or more; three or more such hours in a row; a storm whose peak is ten times baseline or more. |
| uqscan= | A tag family a submitter attached to its urlquery scans. The name is the submitter's, not urlquery's. |
| Cloudflare Worker | A small program run on Cloudflare's edge under a workers.dev name; here, the submitter's own fetch endpoint. |
| canary | A clearly marked instruction placed where only an automated reader would find it, asking for one harmless request so that reading can be measured. |
| honeypot | A disclosed decoy host that looks like an agent console and records who probes it. |
| Shodan | A search engine for internet-connected devices. It scans the internet and indexes what each service returns, such as a web page's title. |
| UA | User agent: the name a client sends with each request. Any client can send any name. |
| ASN | Autonomous system number: the network a request arrives from, which a client cannot choose. |
Sources
- urlquery: hourly scan counts and tags, 27 September to 30 October 2026. Aggregates published with urlquery's approval; individual reports are on their site.
- urlscan.io: daily scan totals for the same series. Aggregates published with urlscan.io's approval.
- Public reporting of the 4 October Amap event: swarmcha.se; TechCrunch, 5 October 2026.
- Shodan: daily host counts and country breakdowns for the listed queries, 21 to 31 October 2026, through the Shodan API. Aggregates only, published with Shodan's permission. © Shodan.
- Descriptions and page titles from the projects' own repositories: ARTEX global edition and Cairn.
- ai.robots.txt: the list of AI crawler names and operators used for labelling (MIT licence).
- Honeypot and canary data: Brown Cyber Services, 22 to 30 October 2026. The honeypot is
agents.browncyber.com; the scanner that checks your own hosts for this kind of exposure is agent-exposure-check.
Corrections and additions: contact@browncyber.com.