Financially motivated intrusion / AI services and crypto exchangesHighTLP:CLEAR

ARTEX AI key theft

ARTEX Agents Steal AI Gateway Keys and Try Deepfake KYC Bypass at Crypto Exchanges

Published
13 November 2026
Added
13 November 2026, 22:30 UK
Last updated
13 November 2026

4 to block now7 to hunt2 behaviours to hunt

Hunt3 platforms

Run the hunt before you block, to find past activity. The same hunt is published for each platform. Preview and copy the queries in the dashboard.

  • Defender XDRKQLDefender portal > Hunting > Advanced hunting. Paste and run over 30 days.Download hunt.kql
  • SplunkSPLSearch & Reporting, last 30 days or more. Reads the CIM data models.Download hunt.spl
  • Google SecOpsYARA-L 2.0Rules Editor. Paste, Retrohunt over 30 days, then turn on live mode.Download hunt.yaral

These queries can return false positives. Know your environment and confirm a hit before acting.

Summary

ARTEX agents steal provider keys from exposed AI gateways and try deepfake KYC bypass at crypto exchanges. Four hosting addresses are on the block list; the Cloudflare, Oracle Cloud and carrier addresses are alert only.

Do now

  1. Block the 4 indicators.
  2. Run the hunt.
  3. Patch Langflow now.

Indicators 11

ValueTypeActionCopy
31[.]22[.]91[.]156IPv4block
216[.]176[.]239[.]231IPv4block
212[.]107[.]28[.]51IPv4block
103[.]142[.]140[.]158IPv4block
104[.]28[.]225[.]223IPv4alert
150[.]136[.]38[.]110IPv4alert
128[.]242[.]115[.]131IPv4alert
104[.]28[.]251[.]98IPv4alert
104[.]28[.]251[.]99IPv4alert
104[.]28[.]219[.]97IPv4alert
125[.]103[.]213[.]41IPv4alert

Sources Amazon Threat Intelligence

More detail
Note Amazon says the operators sent all outbound traffic through proxies, so these addresses are short-lived egress points. Do not block Cloudflare (104.28.x), Oracle Cloud or the NTT and ARTERIA carrier addresses; alert on them only.

What to do

  1. Block the four hosting addresses and alert on the other seven in firewall, WAF, reverse-proxy and cloud sign-in logs for October and November 2026.
  2. Take LLM gateways and AI app builders (One API, New API, LiteLLM, Dify, Langflow, RAGFlow, Open WebUI, Ollama) off the internet or put them behind authentication; change default admin and master keys and finish any unclaimed setup.
  3. Patch Langflow for CVE-2025-3248 and RAGFlow for CVE-2026-28797, and hunt Langflow hosts for shells or downloaders started by the Langflow process.
  4. Rotate every provider API key stored in an exposed gateway, and alert on unexpected AI model spend or calls from new addresses.
  5. Crypto exchanges and KYC teams: review liveness and document checks for injected video and edited ID photos with forged camera metadata.

Behaviours to hunt 2

Behaviours are leads, not confirmed indicators, and can match legitimate activity. Compare any hit against your own baseline and environment before acting.

  1. alert Langflow process starting a shell or downloader

    Look forA shell, downloader or netcat started by the Langflow server process

    Why it mattersAmazon lists Langflow CVE-2025-3248 in the ARTEX playbook; the flaw runs attacker code inside the Langflow Python process, so a payload shows up as a child shell or downloader. Tuning: the parent command line must contain langflow and the child must be sh, bash, dash, curl, wget or nc. Expected false positives: custom Langflow components that shell out by design; baseline those per host before alerting. Amazon Threat Intelligence

    Log sourceDefender XDR DeviceProcessEvents on Linux servers (MDE for Linux); EDR process events in Splunk or Google SecOps

    QueriesKQL · SPL · YARA-L

  2. hunt isCliRequest or default-key use on AI gateway admin APIs

    Look forRequests to One API, New API or LiteLLM admin endpoints from outside your network that carry an isCliRequest header or authenticate with the default master key

    Why it mattersAmazon names the isCliRequest header bypass in One API family systems and the LiteLLM default master key as ARTEX entry points for pulling stored provider keys. Tuning: only external sources, and only management or channel/key endpoints. Expected false positives: your own admin CLI if it sets the header; these logs are not in Defender XDR tables, so no XDR query is given. Amazon Threat Intelligence

    Log sourceWAF, reverse-proxy and AI gateway access logs

    QueriesNone yet: check by hand

Affected countries and sectors

Targets
Worldwide
Sectors
Technology, AI services, Cryptocurrency exchanges, Financial services

Background

Amazon Threat Intelligence reported on 13 November 2026 that its MadPot sensors, posing as LLM inference APIs, captured the working memory of ARTEX agents run by what it assesses, with moderate confidence, are separate Chinese-nexus groups between 3 October and 5 November. The main campaign hunted internet-exposed LLM gateways and AI app builders (One API, New API, Veloera, Sub2API, LiteLLM, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama), pulled the provider API keys they hold and fed them back into the operators' own model pool, then used stolen cloud credentials to reach managed AI services, secrets stores and parameter stores. Exploits included unclaimed installs, anonymous management endpoints, default admin credentials such as the LiteLLM master key, Langflow CVE-2025-3248, RAGFlow CVE-2026-28797, the isCliRequest header bypass in One API family systems and SSRF to cloud metadata. A second campaign tried to pass KYC checks at two cryptocurrency exchanges with real-time deepfake video and forged ID photos; the vendor's server-side check rejected every recorded attempt. Four hosting addresses are on the block list; the Cloudflare, Oracle Cloud and carrier addresses are alert only.

Downloads

View folder in repository

MITRE ATT&CK8

Open in the live dashboard Get the blocklist