Independent · Source-cited · No sponsored rankingsSources linked in every article
AIToolRanked
AI News · 5 min read

MadPot logged ARTEX agents stealing AI keys

From October 3 to November 5, 2026, Amazon's MadPot decoys logged ARTEX agent records of stolen AI keys funding the next search.

RA
· Founder, AIToolRanked
A brass padlock on a closed steel cabinet in a dim server room, with a brass key on top of the cabinet
On this page

What Amazon reported on November 13

On November 13, 2026, CJ Moses published Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents ↗ on the AWS Security Blog. The picture above is ours. We did not run these tools, and we did not replay the traffic.

From October 3 to November 5, 2026, MadPot decoys presented LLM-compatible chat-completions interfaces. Agents sent requests that carried their own operating memory: system prompts, tool histories, knowledge-graph state, and findings. Amazon says that history describes work on systems MadPot does not monitor. Amazon did not watch that work as it happened and cannot verify the outcomes the agents reported.

The largest campaign is theft of AI provider keys. Agents found internet-exposed apps, pulled the upstream keys, checked that they worked, and pooled them. The records show those keys placed in the framework's own gateway, paying for the inference that hunted the next key. Cloud credentials were a door to managed models and secret stores, with premium models ahead of general permissions.

Amazon names twelve products. Gateways: One API, New API, Veloera, Sub2API, and LiteLLM. Builders: Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, and Ollama. The playbook in the records is the list in the next section: claimable installs, open admin endpoints, default keys, two named CVEs, an isCliRequest bypass, metadata SSRF, and secrets left in client files. Discovery used scanners, certificate logs, GitHub, package tarballs, image layers, notebooks, paste sites, and Telegram. Keys public for more than 30 days were graded low-value. A second record set describes deepfake attempts at two cryptocurrency exchanges. Amazon says one vendor server rejected every logged attempt, and Amazon cannot verify outcomes past the records.

Amazon's confidence is moderate, both that the agents ran on ARTEX and that the operators are separate China-nexus groups. It cites prompt references, simplified Chinese, China time zones, targeting, and tooling. It notified the AWS customers it assessed as potentially affected and published 11 IPv4 addresses, first seen from October 14 to November 2, 2026.

Amazon's sentence: "None of the techniques in these campaigns were new. What AI agents change is the speed and persistence of the attacker."

What to lock down

Do these on any host you can reach from the internet. The AWS post names the weaknesses. It does not print a master-key string, a request, or a payload, and this note does not add them.

  1. List every LiteLLM, One API, New API, Veloera, Sub2API, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, and Ollama install, including a weekend proxy on a VPS.

  2. Replace a default administrator credential. If LiteLLM is still on the default master key, set your own. Finish setup, or take the host offline, wherever an uninitialized install can still be claimed. Close management and provider-configuration endpoints that answer without a login.

  3. Upgrade Langflow. The Langflow advisory GHSA-rvqx-wpfh-mfx7 ↗ for CVE-2025-3248 lists every version before 1.3.0 as affected and 1.3.0 as the patched release. It is critical code injection with no login required.

  4. Move RAGFlow off the versions Amazon's CVE points at. The Hacker News ↗ (Swati Khandelwal, November 16) says NVD lists 0.24.0 and earlier for CVE-2026-28797 and that the advisory names no patched version. RAGFlow v0.25.0 ↗, published April 21, 2026, includes "[Security] Fix jinja2 SSTI vulnerability using SandboxedEnvironment." Read that note against the copy you run.

  5. Do not leave a One API-family admin API on the open internet. Amazon names an isCliRequest authentication bypass and names no fix. The Hacker News says the report gives no affected versions.

  6. Block URL-fetch features on these apps from reaching cloud metadata services.

  7. Search client JavaScript, source maps, debug routes, environment files, git history, package tarballs, and image layers for provider keys. Rotate anything that was public. Put a spend limit and an alert on every provider account. Amazon says the stolen keys paid for the next hunt.

What other reports add

Quartz ↗ (Cris Tolomia, updated November 16) reports an FBI FLASH dated November 13: ARTEX in use since at least August, the GitHub repo removed in October, forks and ARTEX PRO still available. Quartz also says Cloudflare saw more than 50 framework families scanning customers in the three days before a November 13 post, and that Cisco Talos credits a July GitHub publication to a developer using the name Autumn-27.

What people are saying

  • "Amazon set up decoy servers that pretend to be LLM APIs, and attackers' AI agents ended up sending their entire campaign memory straight to them" — Haris Nadeem ↗ (@harisn), November 15, 66,284 views when fetched. He adds that he has left LiteLLM and Open WebUI on a VPS without being sure the defaults changed.

  • "Your AI gateway is now a credential vault that attackers scan for, and some of them still run on default keys." — modularciso ↗ (@modularciso), November 14. The profile says a security professional.

  • "The business model is the receipt: a stolen provider key paid for the inference that hunted the next key." — Maksemiilian ↗ (@makmmas), November 16.

  • "the campaign leaked its own memory." — Eugene M. ↗ (@jenistranger), November 16. He calls the sensors honeypots. Amazon's post calls them decoy threat sensors.

Reddit was not reachable for this story.

Our take

If LiteLLM, Dify, Open WebUI, Ollama, Langflow, or RAGFlow is reachable, do the list above first. The records describe default keys, open admin endpoints, secrets in public files, and unpatched software. Amazon's point is how fast that list gets walked.

One open admin panel is every upstream key behind it. The Cursor incident was a vendor key in old traces. The Fireworks breach was a supplier. This record is a key on a host you run. A spend alert is how the stolen key shows up: it paid for the next search.

Ollama on localhost is a different exposure from Ollama on a public port. Local runtimes are in the Ollama and Gemma note. Weights you meant to download are in the open-source LLM comparison. Langflow as a builder is in the agent frameworks comparison. Those reviews do not replace the patch or the rotation.

Do this today if any of the twelve apps has sat on a reachable address, including a forgotten VPS. Teams that only call a hosted model API, and do not run these apps, are outside the list. Amazon cannot verify the deepfake campaign's outcomes. The action in this note is the gateway.

More from the news desk · About Rai Ansar