ARTICLE ORIGINAL MAG212 CYBER

ARTEX Agents Stole AI Keys to Pay for More Attacks: What Amazon Saw and How to Lock Down AI Gateways

novembre 16, 2026 · El Mostafa Ouchen

AI agents that steal keys to pay their way: ARTEX Agents Stole AI Keys to Pay for More Attacks: What Amazon Saw and How to Lock Down AI Gateways. MAG212 CYBER.

De la part de MAG212 CYBER Daily, 16 November 2026. Chaque commande est copiée de la documentation officielle ou de l'analyse citée et a été non testé par MAG212; exécutez-le uniquement sur les systèmes que vous êtes autorisé à administrer.

Statut: Vendor threat intelligence, moderate confidence. Amazon published on 13 November 2026; the findings were widely reported on 16 November. Amazon did not observe the activity live: its view comes from records the agents carried in requests to its decoys, and it cannot verify the outcomes the agents claimed. This is an update to our [15 November ARTEX coverage](https://cyber.mag212.com/fbi-artex-flash-ai-agents-autonomous-intrusions-defender-guide/). [1][4]

Ce qui s'est passé

From 3 October to 5 November 2026, Amazon’s MadPot sensors, configured to look like popular LLM inference APIs, received requests from AI agents. Because agent frameworks send their working history with each call, the requests carried the agents’ own records of their campaigns. Amazon assesses, with moderate confidence, that the agents ran on ARTEX, an open-source agentic penetration-testing framework, and that the operators were separate Chinese-nexus groups, based on simplified-Chinese prompts, China time-zone settings and targeting and tooling patterns. [1]

Campaign 1, AI credential theft. Agents searched for internet-exposed AI applications, extracted the upstream provider keys and cloud credentials they held, validated them and added them to a shared pool fed into ARTEX’s own model gateway. The stolen keys paid for the inference that drove further hunting. The records grade credentials L1 to L3 and treat keys that have been public for more than 30 days as low value. [1]

Campaign 2, identity-check bypass. Agents used real-time deepfake video and forged ID photos against know-your-customer checks at two cryptocurrency exchanges, and forged camera metadata to make edited ID images look unaltered. The liveness-check vendor’s server rejected every recorded deepfake attempt. [1]

Amazon stresses that none of the techniques was new. The agents succeeded where basic hygiene was missing: default or hardcoded credentials, exposed secrets, unauthenticated endpoints and unpatched software. [1]

Qui est concerné ?

  • LLM gateways and proxies: One API, New API, Veloera, Sub2API and LiteLLM. [1][3]
  • AI application builders: Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. [3]
  • Gateways are favoured because a single weak management console can expose many downstream provider accounts. [3]
  • Known vulnerabilities named in reporting include Langflow CVE-2025-3248, fixed in 1.3.0. [3][6]

Pourquoi c'est important

Yesterday's story was the FBI's warning that ARTEX speeds up intrusions. Amazon's data shows a simple business model behind some of it: steal AI keys, spend them on more AI-driven scanning, repeat. Every exposed AI gateway is both a target and, once breached, a source of funding for the attacker. (Analyse MAG212)

AI agents that steal keys to pay their way. Diagram by MAG212 CYBER.
Diagramme : MAG212 CYBER. Rouge plein = observé ou confirmé ; orange pointillé = signalé ou non vérifié ; vert = contrôle recommandé.

Comment fonctionne l'attaque

  1. Scan: agents look for the web interfaces and APIs of the products above.
  2. Accéder: default admin credentials, missing authentication or a known flaw give access to the configuration.
  3. Extract: stored upstream provider keys and cloud credentials are copied.
  4. Validate and grade: keys are tested against providers and graded L1–L3.
  5. Pool and spend: working keys go into a shared pool behind ARTEX’s model gateway and pay for the next round of scanning. [1]

Indicateurs de compromis

IPv4CampaignVu
104.28.225[.]223117–19 Oct
31.22.91[.]156114–19 Oct
150.136.38[.]110131 Oct
128.242.115[.]131131 Oct
216.176.239[.]23112 Nov
212.107.28[.]51214 Oct
103.142.140[.]158217 Oct
104.28.251[.]98 and .99218 Oct
104.28.219[.]97218 Oct
125.103.213[.]41218–19 Oct

Some addresses sit in shared hosting or proxy ranges; check context before blocking. [1]

Ce que les défenseurs doivent surveiller

  • Unexpected spend or usage spikes on AI provider accounts, especially from new regions or at odd hours.
  • New or unknown keys, users or models configured in an AI gateway.
  • Requests to gateway admin endpoints from the IPs above or from unfamiliar networks.
  • Provider keys found in code repositories, container images or public artifacts.

Comment enquêter

Check whether your AI services answer without authentication. Run these from outside your network against hosts you own (read-only; not tested by MAG212):

curl -s -m 5 http://<PUBLIC_IP>:11434/api/tags
curl -s -o /dev/null -w '%{http_code}\n' https://<LITELLM_HOST>/v1/models

Ollama has no built-in authentication, so a JSON model list from the first command means anyone can use it. A LiteLLM proxy configured with a master key should return 401 to the second command; 200 means no key is required. [5][7]

Then review the gateway’s admin and audit logs for logins and key exports, and each AI provider’s usage dashboard for the period from early October.

How to contain and remediate

  1. Put every AI gateway and builder behind authentication and off the public internet. [1]
  2. Change default admin credentials and master keys. Rotate every provider key the tool stored, not only the ones you suspect.
  3. Patch known flaws, for example Langflow 1.3.0 or later for CVE-2025-3248. [6]
  4. Set spend alerts and quotas with each AI provider. [1]
  5. Keep secrets out of code and public artifacts; scan repositories and images.

Comment vérifier la récupération

  • The external checks above fail or return 401.
  • Old provider keys are revoked and fail; new keys are in use only from known systems.
  • Provider usage returns to expected levels.

Prévention et sensibilisation

  • Developers: treat a self-hosted AI gateway like a payment system; it holds credentials worth money.
  • Équipes de sécurité : add AI gateways and builders to external attack-surface scans.
  • Decision-makers: assume exposed secrets will be found within days and fund the basics first, as Amazon advises. [1]

Sujets connexes sur MAG212 CYBER

Sources

  1. [1] AWS Security Blog, “Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents“, 13 novembre 2026.
  2. [3] Tbreak, “AI agents are hunting exposed LiteLLM and Dify tools for API keys“, 16 Nov 2026.
  3. [4] MAG212 CYBER, “FBI ARTEX FLASH: AI agents and autonomous intrusions, a defender's guide“, 15 novembre 2026.
  4. [5] Ollama, “API reference (GET /api/tags)“, docs 2026.
  5. [6] NVD, “CVE-2025-3248 (Langflow, fixed in 1.3.0)“, record 2026.
  6. [7] LiteLLM, “Proxy: virtual keys and master key“, docs 2026.

Written and edited by El Mostafa Ouchen for MAG212 CYBER. Facts checked against the sources above as of 16 November 2026, 6:00 AM PT.