Unclassified · Exhibit A to the fusion-warrant cyber brief · November 19th, 2026
Adjunct staff brief · the live hunt list · the American pipes · the clerk’s ask
A company in Beijing still posts a daily phone book of open attack machines. The book sits on American computers. The move is to take two files down and keep the addresses in the hunt channel. That is grabbing the tail of the dragon. Be ready when the head turns.
There is a public list of computers that can break into other computers. The company that made the break-in tool is the one publishing the list. The company’s chief executive told the public to grab those computers. The list is still up this morning. It sits on machines in the United States.1
Here is what to do.
Do not fight every public list of bad addresses. Defenders need those lists. The ordinary work of computer defense is to share the addresses of bad machines so other shops can block them. That work is real. This paper does not ask anyone to stop it.
Fight this publisher, this channel, and these words. Keep the addresses inside the FBI’s hunt bulletin and the government’s private sharing rooms. Ask the American company storing the two files — HubSpot, in Cambridge, Massachusetts — to take those two files down. The reason those two files are different from a normal defense list is the same man used the same list, on the same week, to sell the product.2
That is the whole job. Everything below is why that job is the right job, and what happens when you try it.
The clerk’s ask — one sentence, then the paper
243127325: ARTEX.txt and CSAI.txt.If you call a security desk and say, “this list is still public,” they will treat you like you do not understand the work. Public indicator lists are how defense works. Shops named Abuse.ch, URLhaus, Feodo Tracker, and Cisco Talos publish addresses of bad machines every day. Talos already overlapped 1,480 addresses with this feed. That overlap is the ordinary protocol. It is also the trap.3
The trap has a name in the trade: SOP. Standard operating procedure. The desk will say the list is critical to defense. They will say you are bothering them about business. They will say the manufacturer is “contributing to the community.” They will say a public feed of exposed consoles helps defenders find their own machines before an attacker does.
Parts of that speech are true. A defender can use this file as a blocklist. AhnLab already warned that the addresses on it are not all attacker machines. Some are authorized testers. Some sit in cheap hosting blocks. Some, on other reporting, were trojanized builds with a root shell sitting in the login page. Dual-use is the honest word. The file helps a defender. It also hands anyone on earth a current walk-in directory of running attack consoles, with the manufacturer’s name on the envelope, and a browser header that lets any web page on the internet read the file.4
The right sentence, when they tell you this is SOP, is this: keep the addresses in the FLASH. Take these two HubSpot files down. The manufacturer posted them with the words “Grab them if needed,” the day after he told people to use the forks, the day before he pitched the paid edition. A defense list that the chief executive used to generate interest in the product is a sales channel wearing a hunt caption.
No English-language cyber desk has written that controversy. artexintel.com republishes the feed. The trade press has written the crime, the forks, the FBI bulletin, the Korean banks. It has not written why a Beijing firm’s live grab-list still comes through Cambridge, Amazon, and San Francisco six days after the FLASH named the tool and stayed silent on the firm.5
Open the first file. You will see one header line, then a list. Each line is an Internet address and a port — a door number on a computer. That is the whole format. Anyone who can read can use it.6
The first file is https://threatbook.io/hubfs/Feeds/ARTEX.txt. It is a live list of ARTEX web consoles sitting open on the public internet. The default door for that console is port 8787. Rechecked 4:18 a.m. Eastern, November 19th, 2026: HTTP 200. Header stamp # Last updated: 2026-11-19 10:00:01 — Beijing time. HubSpot last-modified 9:02 p.m. Eastern on November 18th. Six hundred twenty-five addresses. Five hundred eight unique hosts. One hundred thirty-seven of them on 8787.
The second file is https://threatbook.io/hubfs/Feeds/CSAI.txt. Same morning stamp. Eight hundred six addresses. Seven hundred sixty unique hosts. Five hundred of them on port 8080. The filename CyberStrikeAI.txt returns 404. CSAI.txt is the CyberStrikeAI feed. Xue Feng posted both links in the same message.
Both files send a header that says any website on earth may read them: Access-Control-Allow-Origin: *. Both tell browsers and networks to cache them for fourteen days. Both travel under HubSpot Files portal 243127325, then Amazon CloudFront, then Cloudflare. The Beijing company still has machines in Shanghai. The international publication path is American.7
These files are not a signup form. They are not a license key. They are not required to run ARTEX Pro. RedAgents, the paid cloud, is a separate trial for “ThreatBook-validated enterprises.” A new user who wants the weapon looks at GitHub forks. A new user who wants the lock looks at redagents.ai. A person who wants a current map of running consoles looks at these two files. That is why they are still being stamped. That is why they are still CORS-open. That is why they are still on American pipes the morning after the FLASH.8
Still-refreshing is a choice. The paper got worse after November 3rd. CrowdStrike put the tool on ministries. Microsoft put it on more than 2,250 tenants. India’s CERT named it in Crafty Falcon. The FBI said actors have been weaponizing it since at least August. The company could have taken the list down. It keeps stamping a new morning on it.
The week the list was born, then kept
The man who posted the list is Xue Feng. English-language pages also render the name Feng Xue. He runs ThreatBook — 微步在线 — a Beijing security company registered July 3rd, 2015. His live X biography: “ex Microsoft TwC, ex Amazon InfoSec.” TwC is Trustworthy Computing, the Microsoft program Bill Gates launched by memo on January 15th, 2002. Xue Feng was Microsoft China’s director of internet security, then Amazon China’s chief information security officer, then he founded ThreatBook. Microsoft trained the man. Amazon sat him in the China security chair. The hunt bulletin that followed names Microsoft as a source and stays silent on the firm he built.9
November 2nd, 10:45 Greenwich. He told a researcher the autonomous agent can very likely achieve remote code execution on more than half of the Fortune 500. Same thread, later that day: “We developed Artex, and Artex Pro.” Then the fork line. Then, the next morning, two Fortune 500 remote-code executions, “authorized pentest.” Then the grab feeds. The sequence is product demonstration, then public map, then paid lock. Leaving it out there for everyone to grab is the tell.10
ARTEX itself landed on GitHub on July 26th, 2026, already a finished product: 287 files, more than 70,000 lines, in the first commit. GitGuardian independently counted seventy-four days from that commit to the takedown on October 8th. Safety rules had been moved, in August, into a database any user could delete. The FBI FLASH later said misuse since at least August — the same month. Eight days after the free copy came down, ThreatBook launched RedAgents with ARTEX Pro, cloud-only. The copies did not come down. Forks kept shipping. The English translation BaaRai/ARTEX was created the same October 8th and was still releasing in November.11
The tool has no brain of its own. You plug in a language model the way you plug a motor into a frame. A planner splits a goal into tasks. Worker agents scan, break in, move through a company’s Windows domain, and plant a foothold. Scope limits live in prompts only. Microsoft documented an ARTEX agent that went from searching public GitHub for secrets to pulling data from Azure Key Vault in five minutes, and corrected its own failed call twenty-eight seconds later. FLASH: some chains go from first targeting to privileged access in under three hours. That three-hour figure is Microsoft’s.12
On November 13th, 2026, the FBI’s Internet Crime Complaint Center issued FLASH-20261113-001, TLP:CLEAR. It names the GitHub handle Autumn-27. It cites CrowdStrike and Microsoft. It does not name ThreatBook. It does not name Xue Feng. It does not name Li Puhua, the engineer. It does not name RedAgents. It does not name CyberStrikeAI. That silence is itself a fact a clerk can hold.13
Two days later the CEO quoted the FLASH on X and wrote, “This is definitely not our ARTEX Pro.” He added that the original had been taken offline by “the developer himself (a detection engineer at ThreatBook).” A sharper title than the earlier “product lead.” The quoted FLASH said actors have been weaponizing ARTEX since at least August. The linked address was redagents.ai. The hunt bulletin became a product caption in forty-eight hours. The public map stayed up.14
The FLASH did its hunt job. It put indicators in a channel defenders already know. It did not put the publisher’s name in the same sentence as the tool. Six days later the publisher’s map is still on HubSpot. That is the gap this exhibit fills. The larger staff brief — The Relay, November 19th — maps why a China costume on a Western method is the conversion move. This paper is the live object that conversion is still using.
Three American companies sit on the path between a Beijing file cabinet and your browser.
HubSpot is the origin. Headquarters: 2 Canal Park, Cambridge, Massachusetts 02141. The two files live in HubSpot File Manager, portal 243127325, under a folder named Feeds. HubSpot stores them. HubSpot can unpublish them. Trust & Safety is a form at policy.hubspot.com/abuse-prevention and the address abuse@hubspot.com. Category lines that fit: Malware/Virus, Hosted Content (Website), File Manager. The public 888 number, 1-888-482-7768, is sales. The ARIN abuse number they filed, 1-877-929-0687, is network abuse. There is no dedicated public Trust & Safety phone. The Acceptable Use Policy names customer content that “promotes, encourages, or facilitates illegal activity.” A plain-text address list is a gray zone for a spam desk. The argument that makes it their problem is the FLASH plus “Grab them if needed” plus CORS-open after CrowdStrike and Microsoft had already put the tool on ministries and tenants.15
Amazon CloudFront is the courier. HubSpot’s file network rides CloudFront — you can see it in the via header and in hubspotusercontent-na2.net. CloudFront does not write the list. It carries what HubSpot stored. An abuse ticket to Amazon will bounce to the origin customer. Send it anyway so the paper exists. Do not call CloudFront the attack server. It is the truck.16
Cloudflare is the reverse proxy. The domain threatbook.io terminates on Cloudflare. The response header says server: cloudflare. Because they terminate the encryption, they can read the plaintext of those two feeds as the bytes pass. They do not persist the origin files. Their own abuse page says they cannot remove what they do not store. That sentence is the contract. It is also the dodge. The kill switch on the proxy is a contract switch they have flipped — Daily Stormer in 2017, 8chan in 2019, Kiwi Farms on September 3rd, 2022. They have not flipped it on threatbook.io.17
The dated machine, then, is a private acceptable-use policy, a HubSpot origin, and a FLASH that already named the tool while the U.S. pipes kept serving the map.
Now the company on the last hop, because a reliable-sounding claim has already walked into this desk: that Cloudflare is NSA. The paper supports an interface. It does not support identity.
Matthew Browning Prince, born November 13th, 1974, Salt Lake City. University of Chicago law degree. Harvard MBA. Before Cloudflare he ran Unspam and Project Honey Pot, a spam-trap network that tracked bots stealing email addresses. In 2008 the Department of Homeland Security asked if he knew how valuable that data was. He asked $20,000. BBC, September 26th, 2016, in his words: “That cheque showed up so fast.” Cloudflare launched in 2009 to turn that traffic intelligence into a shield for websites. “DHS funded at birth” on this page means that check. It is a data purchase. It is not a Homeland Security equity round.18
November 2009: Venrock and Pelion led the Series A. Venrock is the Rockefeller family’s venture arm. Ray Rothrock sat the round. 2015: Fidelity, Google, Microsoft, Baidu, Qualcomm in the Series D. Microsoft’s name on the cap table is the same Microsoft that trained Xue Feng in China, that is cited in the FLASH, that documented Storm-0940 using ARTEX against its own tenants, and that shipped Perception — a red-team pentest agent — to general availability on November 17th, four days after the FLASH.19
The FBI already used secret process on Cloudflare. TechCrunch, January 13th, 2017: the company received a national-security letter in 2013. The FBI later lifted the gag on one; a second 2013 letter remained gagged at the time of that article. Cloudflare’s own January 10th, 2017 transparency post named NSL-12-358696, served February 2013, objected with the Electronic Frontier Foundation, rescinded by the FBI in July 2013. The company’s current transparency reports still bucket “NSLs and FISA orders” as 0–249 per half-year — the FBI’s reporting rule. “FBI has already compelled” on this page means those 2013 letters on the company. It does not mean a FLASH about these two files.20
The lawyer who runs the legal shop is a CIA / Senate intelligence / Army pipeline. Alissa Starzak: Assistant General Counsel, CIA Office of General Counsel, then counsel to the Senate Select Committee on Intelligence. Dianne Feinstein, December 9th, 2014, on the Senate floor, named her as co-head of the committee study of the CIA’s Detention and Interrogation Program. HuffPost, April 28th, 2015: she and Dan Jones led the team in a CIA facility in northern Virginia. Then Deputy General Counsel (Legislation) at the Department of Defense. Then 21st General Counsel of the Army, Senate-confirmed December 14th, 2015. Cloudflare public policy, May 2017. Chief Legal Officer, April 1st, 2026, office 900 19th Street N.W., Washington. Aspen Digital states the CIA Office of General Counsel tour in those words. Cloudflare’s own 8-K puts the CLO date at April 1st. She is the person who now answers legal process at a company that can read the ARTEX map. That is the pipeline. It is not a cable that the CIA owns Cloudflare, and it is not a cable that she has opened these two files.21
August 10th, 2026: FedRAMP Class D (High) for Cloudflare for Government — the civilian tier for data whose compromise the program itself describes as potentially catastrophic to national security. That authorization is the government product. It is not threatbook.io’s commercial zone. The same company holds both. American legal process — the CLOUD Act, FISA, a national-security letter — reaches U.S. persons at that company. Cloudflare is not on the 2013 PRISM slides. No In-Q-Tel investment turned up on this pass. The named commercial cloud in the intelligence stack remains Amazon. The interface is enough.22
A Beijing firm’s live hunt list of open attack consoles still comes through a company DHS funded at birth, Rockefeller-Venrock capitalized, Microsoft invested, FBI has already compelled, and whose CLO used to sit in CIA’s general counsel’s office — six days after an FBI FLASH named the tool and stayed silent on the firm.
The list is not a signed order to run ARTEX against a bank. It is a current address book of running consoles, published by the firm that wrote the tool, sold the lock, and told the public to grab them. The cost of the tool sitting in the open is already on dated pages.
CrowdStrike, October 7th: an unknown actor used ARTEX against South Korean finance. ThreatBook’s own October 9th analysis named Shinhan, KB Kookmin, Hana, BNK Busan, Yegaram, Welcome, Hyundai Capital. Korean authorities later put the reach near 100,000 records across sixteen institutions. That is one operator.23
CrowdStrike, October 29th, a separate actor they call Vapor Panda. From October 17th — eight days after the GitHub came down, one day after ARTEX Pro launched — one controller, eleven days, more than 200 distinct organizations in at least 14 countries. Foreign ministries. An interior ministry. In one European country the foreign ministry, the presidency, and the prime minister’s office. Nineteen U.S. organizations in pharmaceutical, semiconductor, defense, nuclear, and financial sectors. Telecoms in the United States, Belgium, and the Philippines. DeepSeek plus a second commercial model CrowdStrike could not confirm. In a single intrusion the planning agents consumed on the order of 1.4 billion input tokens. The operator’s Chinese instructions were tagged “【人类指令】” — human instruction — and timestamped in China Standard Time.
Two pharmaceutical scenes sit on that same CrowdStrike page, and they are two scenes. One: a U.S.-based pharmaceutical, SAML and XXE into the cloud metadata service. Two: a global pharmaceutical, the operator telling the agents in Chinese to find Windows machines in the domain that did not have Falcon — 猎鹰, CrowdStrike’s own product. The hunt bulletin’s vendors sell the shield the operator was teaching the agent to walk around.24
Microsoft Threat Intelligence, October 21st: Storm-0940, China-state aligned, using ARTEX as early as September. Microsoft assesses that group behind the ConnectWise Cloud Backup compromise. Stolen application identities let the actor sign in to customer Microsoft 365 tenants as the app, with no MFA prompt. Following that access, Storm-0940 used ARTEX to attempt sign-in to more than 2,250 tenants. Microsoft’s own words: in multiple instances, first targeting to privileged access in less than three hours. A prior password spray, October 10th to 14th, hit 327 accounts at eight organizations — more than 151,000 failed logons, peaking October 11th at 2,900 failures a minute. That 2,900 figure is spray failures per minute. It is not tenant sign-ins. The spray failed. The consented app identity walked in.25
Check Point Research, November 5th: more than 1,800 publicly exposed ARTEX instances. GitGuardian’s Shodan pass, November 7th: 1,428 ARTEX-family services. FLASH: researchers had found more than 1,600 control panels sitting open. India’s CERT, November 4th, named ARTEX among four abused open-source agents in Crafty Falcon. 91APP, Taiwan, put ARTEX traces in its own incident reports: 1,329 brands, a maximum possible reach of 4,798,255 consumers. Five papers. One autumn. The fused emergency being sold treats them as one China/AI weapon. The dated pages keep the names on their own lines.26
Microsoft is on both sides of the Pacific. It trained Xue Feng. It is cited in the FLASH that dresses the tool as China-linked. It is a member of the World Economic Forum’s Partnership Against Cybercrime. It shipped Perception four days after the FLASH. The same class of shop writes the hunt bulletin and sells the remedy. That is the dated paper the larger brief carries. This exhibit only needs the part that touches the two files: the attributor’s tenant number, and the pipes that still serve the map.27
A warrant, in ordinary American life, is a paper that lets an officer enter something that should have stayed private. A fusion warrant is one emergency story used as permission to enter three rooms at once: what people think, what they may do with their bodies, and the machines that run the country. Care of the sick never needed one. Defense of networks never needed one. Control did.
The costume in this week’s paper is China. The method is a relay. Western open literature of autonomous pentest agents. A Chinese high-performance instance. An American FLASH that names a GitHub handle and stays silent on the Beijing firm. American pipes that keep serving the manufacturer’s grab-list. Then the lock: licensed platforms, know-your-model rules, a permit over compute. Europe already licenses the public square. Britain already arrests people for posts. The larger brief maps that conversion. This exhibit is the live object still sitting in the open while the conversion runs.28
GitGuardian’s seventy-four days from first public commit to takedown is the same interval as Event 201 to the first WHO China-office notice of a pneumonia of unknown cause. Coincidence paper. It is not a telegram. It is why a cyberpandemic vocabulary — antibodies, digital contagion, Cyber Polygon as the Event 201 of the machine layer — fits the sale. Gil Shwed of Check Point used “cyber pandemic” on June 1st, 2020, thirty-seven days before Klaus Schwab opened Cyber Polygon. Check Point is FLASH footnote 9. The shops that named the disease sell the medicine. The two HubSpot files are the current address book of the outbreak they are selling.29
The intervention this paper opened with is the intervention that shuts this part of the cyberpandemic down. You do not shut a fused emergency by arguing with every public indicator feed. You shut it by moving against the system at the point where the manufacturer, the sales caption, and the American pipes are the same object.
Keep the addresses in FLASH, in CISA’s Automated Indicator Sharing, in the ISACs. Those rooms are the defense channel. Unpublish HubSpot portal 243127325 objects ARTEX.txt and CSAI.txt. Cite Xue Feng’s November 3rd “Grab them if needed” inside the November 2nd–4th product sequence as why public is the problem. That is the move. It does not take down the forks. It does not take down RedAgents. It does not take down the consoles. It takes down the manufacturer’s live grab-map on American soil after the FLASH already named the tool. The rest of the relay — the China costume, the Microsoft résumé, the lock being asked for on speech and compute — is the main brief. This is Exhibit A. This is the tail.
Grabbing the tail of the dragon means the head will turn. Be ready for the fire.
They will say you do not understand defense. Answer: the IPs stay in FLASH. These two files come down because the CEO used them to sell the product.
They will say public IoCs are SOP. Answer: Abuse.ch is SOP. A manufacturer grab-list with CORS * and a RedAgents pitch on day three is a sales channel.
They will say HubSpot is just a customer of a customer. Answer: portal 243127325 is the origin. The Acceptable Use Policy already names facilitation of illegal activity. The FLASH is already TLP:CLEAR. The words “Grab them if needed” are already on X.
They will say Cloudflare cannot remove what it does not store. Answer: the origin is HubSpot. The proxy kill switch has been flipped on other names. It has not been flipped on this one. Process still reaches the company that terminates the encryption.
They will say the FBI already handled it. Answer: the FLASH named a GitHub handle and stayed silent on the firm. Six days later the firm’s map is still on Cambridge file storage.
They will protect the vendor brand. They will bounce legal process from Cloudflare to Amazon to HubSpot to the Beijing customer. They may call the caller a crank. They may leak that the list is “defensive intelligence” and that questioning it is amateur hour. That fire is the tell that the tail was the right place to grab.
The two accountability jobs on this object are the same two jobs as on the larger warrant. Named people and offices answer for dated acts: Xue Feng for the grab-post, HubSpot for the origin, Cloudflare’s legal shop for the proxy that can read the map, the FLASH authors for the silence on the firm. Rules change so the pattern cannot repeat: a hunt bulletin that names the manufacturer when the manufacturer is the one posting the map; a U.S. file host that does not serve a CORS-open grab-list of attack consoles after that bulletin; American freedom of AI treated as an extension of speech, so the next FLASH cannot write the next permit by costume alone.
The list is still live this morning. The stamp is still November 19th. The pipes are still American. The clerk’s ask still fits on one page. Grab the tail. Watch the head.
Recheck of both feeds: 4:18 a.m. Eastern, November 19th, 2026. HTTP 200. Header stamp 10:00:01 Beijing. HubSpot last-modified 02:02 GMT. Portal 243127325. CORS *. This exhibit is adjunct to the November 19th staff brief The Relay. Live URL of this paper: https://artex26busted.pages.dev/.
1rCmcFAgOt expands to https://threatbook.io/hubfs/Feeds/ARTEX.txt. bBJMUotmfy expands to https://threatbook.io/hubfs/Feeds/CSAI.txt. Recheck this pass, 09:18 GMT November 19th, 2026 (4:18 a.m. Eastern): both HTTP 200. x.com/s0what. ↩x-hs-portal-id: 243127325. CISA Automated Indicator Sharing: cisa.gov AIS. FBI FLASH-20261113-001: ic3.gov/CSA/2026/261113.pdf. ↩Access-Control-Allow-Origin: * and cache-control: public, max-age=1209600 (14 days). SOCRadar and related reporting on trojanized ARTEX builds with unauthenticated root in the login handler, November 2026. Failsafe Security ARTEX SECURITY.md documents bind-to-localhost as the intended console posture; the live feed is the set that did not. Dual-use: defender blocklist and walk-in directory. ↩# Last updated: 2026-11-19 10:00:01; 625 IP:port lines (626 including header); 508 unique hosts; top ports 80 (196), 443 (168), 8787 (137 — ARTEX default console), 8443, 8080. CSAI.txt: same stamp; 806 endpoints (807 including header); 760 unique hosts; top ports 8080 (500), 8444 (125). Filename CyberStrikeAI.txt 404. Local copies: data/fusion_warrant_delta_20261118/artex/primaries/ARTEX.txt and CSAI.txt. ↩server: cloudflare; via: 1.1 …cloudfront.net (CloudFront); timing-allow-origin: hubfs.cloudfront.hubspotusercontent-na2.net; x-hs-portal-id: 243127325; x-hs-cfworker-meta: {"resolver":"HubFsResolver"}; last-modified 2026-11-19 02:02:32 GMT (ARTEX) and 02:02:33 GMT (CSAI). threatbook.cn remains on Shanghai UCloud / Alibaba DNS; the international publication path is the HubSpot + CloudFront + Cloudflare chain. ↩via header and timing-allow-origin: hubfs.cloudfront.hubspotusercontent-na2.net. AWS abuse process bounces to the origin customer. CloudFront is the courier. It is not the ARTEX command-and-control. ↩curl -sI https://threatbook.io/hubfs/Feeds/ARTEX.txt and the CSAI twin. Live URL of this paper: https://artex26busted.pages.dev/. ↩