Threat feed

ARTEX self-funding AI key theft

ARTEX agents steal AI model keys from exposed gateways and spend them on further autonomous attacks

Posted · Spectrum Delta Research Team

Summary

Chinese-nexus threat groups ran autonomous AI agents on the ARTEX penetration-testing framework that stole model-provider API keys from exposed AI gateways and used the stolen keys to pay for the inference powering further key hunting, Amazon Threat Intelligence reports (B/2). Between October 3 and November 5, Amazon's MadPot decoy sensors posing as LLM inference APIs captured the agents' own operating memory, revealing a self-sustaining loop in which working keys were pooled and fed into the framework's model gateway. Targets included the gateways One API, New API, Veloera, Sub2API and LiteLLM and the app builders Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama, exploited through default credentials, uninitialized installs and known flaws. A parallel campaign used real-time deepfake video at 11.5 frames per second and forged ID photos to attempt know-your-customer bypass at two cryptocurrency exchanges; every recorded liveness attempt was rejected. Material uncertainty: Amazon assessed ARTEX use and Chinese-nexus involvement with moderate confidence and could not verify outcomes the agents reported, since most activity occurred on systems MadPot did not monitor.

Breakdown

Targets

  • Industry: Self-hosted AI gateway operators
  • Industry: Cryptocurrency exchanges

Technical exposure

  • Internet-exposed LLM gateway or AI app builder
  • Default credentials or unauthenticated admin endpoints
  • Provider API keys stored in app configuration
  • No monitoring of LLM API spend anomalies

Adversary

Unattributed Chinese-nexus groups · Amazon Threat Intelligence, moderate confidence (B/2)

Procedures (5)

  1. Agents reconnoiter exposed AI applications at scale

    • T1595.002Active Scanning: Vulnerability Scanning

    Logs

    • WAF or CDN logs · Systematic scanning of management and provider-configuration interfaces on LiteLLM, One API, Dify and similar deployments; discovery pivoting through internet scanning services, certificate transparency, GitHub repositories, npm and PyPI tarballs, container-image layers and paste sites
    • GitHub audit logs · Secret-scanning alerts followed within hours by validation probes against the exposed keys from proxy networks
  2. Known attack chains exploited on AI platforms

    • T1190Exploit Public-Facing Application
    • T1078.001Valid Accounts: Default Accounts

    Logs

    • Web application logs · Admin access using LiteLLM default master key; isCliRequest header authentication bypass on One API family systems; exploitation of Langflow CVE-2025-3248 and RAGFlow CVE-2026-28797; anonymous access to provider-configuration endpoints
    • Container and host logs · Uninitialized installations answering on management ports; SSRF probes against URL-fetch features and cloud metadata services
  3. Provider keys extracted, validated and pooled

    • T1552.001Unsecured Credentials: Credentials In Files

    Logs

    • LLM provider usage and billing logs · API calls with valid customer keys from unfamiliar source IPs or proxy networks; usage and spend spikes on accounts whose owners report no such activity
    • Cloud audit logs · Enumeration of storage buckets, secrets stores, parameter stores and managed AI model services using validated cloud credentials, prioritizing premium frontier-model access
  4. Stolen keys fund further autonomous hunting

    Logs

    • LLM provider audit logs · Victims' keys used through the ARTEX model gateway for inference at anomalous hours, from regions outside the customer base, or at volumes inconsistent with the account; credentials graded by working calls and spendable balance, with keys exposed over 30 days treated as low value

    Detectors

    • Amazon MadPot · Decoy LLM inference API sensors capture agent requests carrying ARTEX campaign memory
  5. Deepfake video attempts exchange KYC liveness checks

    Logs

    • KYC platform session logs · Failed liveness sessions driven by a face within 7 pixels of the target position; ID photos with synthetic camera metadata and preserved original compression tables; multiple simultaneous sessions researching cash-out methods while clearing balances

Indicators of compromise

IPs (11)

  • 104.28.225.223Campaign 1 credential theft, seen 2026-10-17 to 2026-10-19
  • 31.22.91.156Campaign 1 credential theft, seen 2026-10-14 to 2026-10-19
  • 150.136.38.110Campaign 1 credential theft, 2026-10-31
  • 128.242.115.131Campaign 1 credential theft, 2026-10-31
  • 216.176.239.231Campaign 1 credential theft, 2026-11-02
  • 212.107.28.51Campaign 2 KYC bypass, 2026-10-14
  • 103.142.140.158Campaign 2 KYC bypass, 2026-10-17
  • 104.28.251.98Campaign 2 KYC bypass, 2026-10-18
  • 104.28.251.99Campaign 2 KYC bypass, 2026-10-18
  • 104.28.219.97Campaign 2 KYC bypass, 2026-10-18
  • 125.103.213.41Campaign 2 KYC bypass, 2026-10-18 to 2026-10-19

CVEs (2)

  • CVE-2025-3248Langflow vulnerability in the agents' codified exploitation playbook
  • CVE-2026-28797RAGFlow vulnerability in the agents' codified exploitation playbook

Sources

Questions, detections or sightings of ARTEX self-funding AI key theft? Bring them to the community.