Back
17/11/2026
The AI agents that steal keys to pay for themselves: inside the ARTEX campaigns
Amazon’s decoy AI APIs caught attacking agents’ own notes: stolen AI keys paying for the hunt for more keys. What happened, and what to fix.
Amazon’s security team has published a rare kind of threat report. Its decoy servers, built to look like AI model APIs, were used by attacking AI agents as their own “brain”, and the agents sent along their campaign notes with every request. For once, defenders could read the attacker’s working memory. What it shows is not a futuristic AI attack. It is old hygiene failures, found and exploited faster and more cheaply, with stolen AI keys paying for the hunt.
The trap: decoy AI APIs that read the attackers’ notes
Amazon runs a large network of decoy sensors called MadPot. Some of them pose as popular large language model (LLM) inference APIs, the services an application calls to get answers from a model. Between 3 October and 5 November 2026, AI agents connected to these decoys as if they were a real model provider.
An AI agent has to send its context to the model on every call. So the decoys received the agents’ system instructions, tool histories, knowledge-graph state and reported findings. In effect, the attackers mailed their playbook to Amazon.
According to Amazon, the agents ran on ARTEX, an open-source agentic AI penetration testing framework. Amazon assesses, with moderate confidence, that the campaigns were run by several separate China-nexus groups. Its evidence includes prompts written in simplified Chinese, China time-zone settings, and “red lines” in the instructions that excluded targets in mainland China and government domains. Moderate confidence is not proof, and the attribution should be read that way.
Amazon is also clear about the limits. It did not watch the attacks happen. It saw only what the agents sent to its sensors, and it cannot verify the outcomes the agents claimed.
Campaign 1: stolen keys that pay for the next theft
The largest campaign hunted for credentials to AI services: the API keys companies use to pay for commercial AI models. The agents went after two kinds of self-hosted software.
- LLM gateways and proxies: One API, New API, Veloera, Sub2API and LiteLLM. These sit in front of several AI providers and hold all their keys in one place.
- AI application builders and chat front ends: Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama.
The techniques were not new. The agents looked for:
- fresh installs that nobody had claimed yet, where the first visitor becomes the administrator;
- admin and provider-configuration pages that answered without a login;
- default administrator credentials, such as LiteLLM’s default master key;
- known vulnerabilities, including Langflow CVE-2025-3248 and RAGFlow CVE-2026-28797;
- an authentication bypass through the isCliRequest header in the One API family of gateways;
- server-side request forgery (SSRF) to reach cloud metadata services and their credentials;
- secrets left in client-side JavaScript, source maps, debug routes and environment files.
The agents graded what they found: a lead, then a key confirmed with a working API call, then a key with measured impact. Keys that had been public for more than 30 days were treated as low value, on the assumption that someone else was already using them.
The part that should worry every company running AI services is the economics. Working keys were fed back into the attackers’ own model gateway, so the stolen keys paid for the AI inference that powered the search for more keys. Amazon describes this as a self-sustaining operation. The victims footed the bill, and the attack cost the attackers almost nothing to keep running.
Campaign 2: deepfakes against identity checks
The second campaign used the same kind of agents against know-your-customer (KYC) checks at two cryptocurrency exchanges. The agents produced forged ID document photos with fake camera metadata, and real-time deepfake video to pass live selfie checks. One session built a real-time deepfake pipeline on a single consumer graphics card, running at 11.5 frames per second.
It did not work. In every attempt the agents recorded, the identity vendor’s server-side check rejected the deepfake, even when the best attempt kept the face within 7 pixels of the target. The lesson is that the final verdict must come from the server, not from anything the user’s device can control.
Why AI gateways are the crown jewels
An LLM gateway is convenient because it puts every AI provider behind one admin page. That is also why it is the best target. One weak login, one unclaimed install or one exposed configuration endpoint gives away every provider account behind it at once, along with their spending limits.
Many of these tools are installed quickly for a pilot or a hackathon and then left running on the internet. They often sit outside the normal patching, monitoring and access reviews that cover production systems. ARTEX-style agents scan for exactly that gap, all day, at very low cost.
What to do now
- Treat AI gateways and AI app builders as production systems. Put them behind authentication and, where possible, a VPN or private network. Include them in patching, monitoring and access reviews.
- Finish every install. Claim the admin account immediately after installation, and remove default passwords and master keys, including LiteLLM’s default master key.
- Close the admin interfaces. Admin and provider-configuration pages should never answer without a login, and should not be reachable from the internet.
- Patch the named vulnerabilities. Langflow CVE-2025-3248 is fixed in Langflow 1.3.0 and is on CISA’s list of actively exploited vulnerabilities. For RAGFlow CVE-2026-28797, the advisory lists no patched version, but from RAGFlow 0.25.0 the affected components render templates in Jinja2’s sandbox, so upgrade to 0.25.0 or later. We found no published fix for the One API isCliRequest bypass, so keep One API and its forks off the public internet.
- Rotate provider keys. If a gateway or AI app was ever exposed, assume its keys were taken. Rotate them at each provider, not only in the gateway.
- Set spending limits and alerts. Give each key its own budget cap and alert on unusual spend or usage. A sudden bill is often the first sign of a stolen key.
- Keep secrets out of what you ship. Check front-end JavaScript, source maps, debug routes and environment files for keys before every release.
- Block metadata abuse. On cloud servers, require the newer metadata service versions that resist SSRF (for example IMDSv2 on AWS), and limit what the instance role can do.
- Check the indicators. Amazon’s report lists IP addresses seen in both campaigns. Search your gateway and web logs for them.
The bottom line
ARTEX did not invent a new attack. It took default passwords, open admin pages and unpatched software, and made exploiting them fast, constant and self-funding. The defence is the same basic hygiene as always, applied now to AI gateways and AI tools as well. Assume any exposed key or misconfiguration will be found quickly, and fix it before it is.
Want help reviewing how your AI tools and gateways are exposed? Contact the LOGIQ team.
Sources
- AWS Security Blog: Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents (13 November 2026)
- NVD: CVE-2025-3248, Langflow unauthenticated remote code execution (fixed in 1.3.0)
- RAGFlow security advisory GHSA-vvwj-fvwh-4whx: CVE-2026-28797, server-side template injection
- CISA: Known Exploited Vulnerabilities catalog
Photo: Lucas Seebacher / Pexels
Japan
Korea
China
Taiwan
Vietnam
Thailand
Indonesia
Portugal
Spain
France
Germany
Egypt
Russia