Illustration of a white server with its door open and a blue key sticking out, as a purple magnifying glass moves in to find it

AI Agents Hunt Exposed AI Apps to Steal Keys: What to Check

A
Admin
·

On 13 November 2026, Amazon reported campaigns in which attacker-run AI agents scan the internet for exposed AI applications, such as LiteLLM, Dify, Open WebUI and Ollama, to pull out the model-provider keys stored inside. If your business self-hosts one of these tools on a server, this applies to you.

What we know

  • Who reported it: Amazon's threat intelligence team, on the AWS Security Blog, on 13 November 2026. It caught the activity with MadPot, a network of decoy servers posing as AI model APIs, between 3 October and 5 November 2026.
  • What the agents did: according to Amazon, they found AI apps and gateways open to the internet, extracted the provider keys stored in them, checked that the keys worked and added them to a shared pool. They favored apps that hold keys for several providers.
  • Which software they targeted: Amazon names One API, New API, Veloera, Sub2API, LiteLLM, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama.
  • How they got in: default credentials (Amazon cites LiteLLM's default master key), unfinished installs anyone could claim, authentication bypasses, server-side request forgery against cloud metadata services, and secrets left in client-side JavaScript, source maps, debug routes and environment files. Also known flaws in Langflow (CVE-2025-3248) and RAGFlow (CVE-2026-28797).
  • Where they looked for keys: per Amazon, internet scanners, certificate transparency logs, GitHub repositories, npm and PyPI packages, container image layers, paste sites and Telegram channels.
  • When they found cloud credentials, the agents went after that account's managed AI model services, again according to Amazon.
  • Who is behind it: Amazon has moderate confidence that the agents ran on ARTEX, which it describes as an open-source agentic AI penetration testing framework, and that the operators were separate China-linked groups.
  • Scale: the agents' records covered "hundreds of targets," Amazon says. It notified AWS customers who may have been affected and published 11 IP addresses as indicators.
  • What Amazon can't confirm: it says it could not verify the outcomes the agents reported, since it did not observe that activity as it happened.
  • A second campaign: Amazon also logged agents trying to beat selfie-based identity checks with real-time deepfake video; the verification vendor rejected it in every recorded attempt.

What changes and what doesn't

  • No new technique. "None of the techniques in these campaigns were new," Amazon says. What is new is scale: an agent probes hundreds of servers without a break and finds in hours what used to take weeks.
  • A small slip costs more. A stolen AI key is cash for the attacker and a bill for the owner. One poorly locked gateway can expose keys for several providers at once.
  • Not affected: people who use ChatGPT, Gemini or Claude through the official website or app, and ordinary WordPress or Drupal sites with no AI tools installed on their server.
  • Affected: anyone who installed, or had someone install, one of these apps on a VPS, a cloud server or an office machine reachable from the internet, even "just for testing."

How to tell if this affects you

  1. Make a list. Write down which AI apps run on servers you control: the chatbot on your site, a model gateway such as LiteLLM, a Dify or Open WebUI for your team, a test Ollama. Test setups and anything a vendor set up for you count too.
  2. Check what the outside world sees. Open each app's address from your phone on mobile data, off the office network. If you get a login screen, it is exposed; if you get the dashboard with no login, shut it now. The safer setup is behind a VPN, an IP allowlist or a proper login.
  3. Change the defaults. Admin passwords and, in LiteLLM, the master key. If an install was left half-finished, finish it or turn it off: an unfinished installer can be claimed by anyone.
  4. Hunt for stray keys. Make sure no AI key shows up in code your visitors' browsers download (in the browser, "View source" and search for "sk-" or "key"), in .env files reachable from the web, in public repositories or in published container images. If your site has a chatbot, the key belongs on the server, never in the JavaScript.
  5. Update. Langflow and RAGFlow appear in the report with specific flaws: keep them on their latest release and turn off open user sign-up if you don't need it.
  6. Set caps and watch usage. With each model provider (OpenAI, Anthropic, Google and so on), turn on a spending limit and alerts, and review the last few days of usage. A spike your traffic doesn't explain is the fastest signal.
  7. If something was exposed, treat its keys as stolen: create new keys, swap them into your apps and revoke the old ones. Do the same with any cloud credentials on that server.

Related: LLMjacking: Stolen AI Keys Run Up Cloud Bills

Sources

Updates: any changes backed by a primary source will be added here.