Guide · AI security · as of 17 November 2026

AI agents are stealing AI keys to pay for their own attacks. Five checks.

On 13 November 2026 Amazon’s security team published what it saw between 3 October and 5 November: attack agents built on ARTEX sent their own working notes to Amazon’s decoy servers. The notes showed what the agents had been told to do. The biggest job was stealing the keys that let software use paid AI models. An AI key is a password that lets a program use a model such as Claude or GPT and bills the owner for it. This guide works from Amazon’s report and turns it into five checks for any business that has added AI to a website, an app or an internal tool.

What Amazon saw

Stolen keys paid for the next theft.

Amazon runs decoy servers that look like real services. Some were set up to look like AI model services. The attack agents treated them as a place to do their thinking and sent them their full task history. Amazon calls the main campaign “a self-sustaining operation in which stolen keys paid for the LLM inference powering further credential hunting.” In plain words: each stolen key bought the AI time used to find the next one. Amazon is careful about the limits, and so are we. It read the agents’ own records, did not watch the break-ins happen, and says it cannot verify the results the agents reported. The records name hundreds of targets, and Amazon says it told the customers of its own that were mentioned. A second campaign tried to pass identity checks at two cryptocurrency exchanges with fake video; every attempt the agents recorded was rejected.

The FBI’s alert on ARTEX, and four checks

Where the keys were

One tool, many keys.

The agents went for software that holds several AI keys in one place, because, as Amazon puts it, “one weak management interface can expose many downstream accounts.” It names twelve. Gateways that pass requests on to model providers: One API, New API, Veloera, Sub2API and LiteLLM. And tools for building AI apps and chat screens: Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. The ways in were ordinary. A setup screen nobody finished, so the first visitor becomes the administrator. A settings page that answers without a login. An administrator password still at its factory value; Amazon names the LiteLLM default master key. Two known flaws left unpatched, CVE-2025-3248 in Langflow and CVE-2026-28797 in RAGFlow. And keys left where a browser can fetch them: in a site’s JavaScript, its source maps, a debug page or an environment file.

Your AI server trusts localhost: four checks

Five checks from Amazon’s report

Run these this week.

  • 1. List every tool that holds an AI key: gateways, app builders, chat screens, model servers, and the test copy someone set up in the spring. Test: for each one, write down who owns it and whether it opens from outside the office network. If it opens and nobody owns it, switch it off first.
  • 2. Finish the setup and change every default. Test: open each tool’s address in a private browser window. You should get a sign-in page, not a setup screen and not a settings page. Then try the default administrator password or master key from the product’s own documentation and confirm the tool refuses it.
  • 3. Patch the named flaws. Test: compare the version you run with the fixed version in the vendor’s notice for CVE-2025-3248 (Langflow) and CVE-2026-28797 (RAGFlow). If you run a gateway from the One API family, update it; Amazon reports a way past its login.
  • 4. Keep keys out of anything a browser can fetch. Test: from outside, ask your site for /.env and /.git/config; both must return “not found”. Have your developer search the site’s JavaScript and source maps for key patterns. A key that was ever public gets replaced, not hidden.
  • 5. Give every AI key a spending limit and an alert. Amazon’s advice is to watch for unexpected spend or unusual usage. Test: find the limit on each key in the provider’s console and name the person who receives the alert. No limit and no name means a stolen key is found when the invoice arrives.

What we checked on our own

Thirteen sites, read from outside.

On 17 November 2026 we ran check four from outside on 13 sites and apps we run. We asked each one for six files that should never be public, such as an environment file and a Git settings file: 78 requests, and no real file came back. We read the 135 JavaScript files their front pages load and searched them for AI, cloud and service key patterns: none found, and no open source maps. It is a narrow check and we say so. One more app sits behind a password and was not read. It covers the files a front page loads, not everything an app fetches later, and it did not look at spending limits on our own AI accounts. Of the twelve tools Amazon names we run one, Ollama, on an office workstation with no public address; its log since 1 October shows calls from that machine only, and none of the addresses Amazon published.

AI agents in Dubai, built with the keys kept on the server

Questions

Questions about stolen AI keys.

  • What is an AI API key?

    A long password that lets a program use an AI model and charges the account that owns it. Anyone who holds the key can run up that account’s bill and, on some services, read what is stored with it.

  • Why would anyone steal an AI key?

    To use expensive AI models while someone else pays. In the campaign Amazon describes, the stolen keys were fed back into the attackers’ own agents, so each theft paid for the search for the next.

  • How do I know if my AI key was stolen?

    Look at the usage page of each AI provider: spending you cannot explain, use at hours nobody works, or models your product never calls. Amazon’s report also lists the internet addresses it saw, which a developer can compare with your server logs.

  • What is an LLM gateway?

    Software that sits between your apps and the AI providers and holds the keys for all of them, so each app does not need its own. LiteLLM and One API are examples. That convenience is why attackers look for them: one weak login exposes every key inside.

  • What should I do if a key was exposed?

    Replace it at the provider, then find how it got out. Deleting the file or the line of code is not enough, because copies stay in old versions and caches. Amazon’s report says the agents treated keys public for more than 30 days as low value, which means the first days matter most.

  • Does this affect businesses in the UAE?

    Amazon’s report does not name countries or victims. It says the agents worked through a fixed list of countries and went for company production sites first. The tools it names are used here as much as anywhere, so the checks are the same.

Sources

read 17 November 2026

Check it at the source.

  1. Amazon Threat Intelligence: threat actors running cybercrime campaigns with ARTEX AI agents (the full report of 13 November 2026: what the decoys captured, the twelve tools, the ways in, and the addresses seen)
  2. FBI FLASH-20261113-001: Threat Overview of ARTEX Autonomous Penetration Testing Framework (the FBI’s alert on the same tool, published the same day)
  3. BRev Digital: Your AI coding tool may be keeping your passwords in plain text (another place keys collect: the settings files of AI coding tools)

A question?

Not sure where your AI keys live?

Tell us which AI tools your business uses, even roughly. We will help you list where each key sits, run these five checks with you, and say plainly what to close first.

Talk to usWhatsApp +971 50 728 5494hello@brevdigital.com