Amazon caught AI agents hunting exposed LLM gateways for API keys. Lock yours down.
SummaryAmazon's decoy sensors captured attacker AI agents' own campaign records: they scanned for exposed LiteLLM, One API, Dify and similar apps, pulled the provider keys inside, and used stolen keys to pay for more hunting. Every technique was old. Here's the checklist.
· MODEL Editors · 2 min read
Some links may be affiliate links. Disclosure

What happened. From Oct. 3 to Nov. 5, 2026, Amazon's MadPot network of decoy sensors posed as popular LLM chat-completions APIs. Attacker AI agents sent prompts to them as if they were a model backend, and those prompts carried the agents' own campaign memory: instructions, tool histories and reported findings. Amazon published what it saw on Nov. 13 (AWS Security Blog (opens in a new tab); The Hacker News (opens in a new tab); Quartz (opens in a new tab)).
The main campaign: stealing AI keys. According to the agents' records:
- They looked for internet-exposed AI apps, pulled out the upstream provider keys stored inside, checked the keys worked, and added them to a shared pool.
- Stolen keys were fed into the framework's own model gateway and paid for the agents hunting for more keys.
- They favored apps that hold keys for many providers, because one weak admin page can expose many accounts.
- Named targets: LLM gateways and proxies One API, New API, Veloera, Sub2API and LiteLLM; app builders Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama.
- They also searched GitHub, npm and PyPI tarballs, container image layers, notebooks, paste sites and Telegram channels for leaked keys, and treated keys public for more than 30 days as low value.
The playbook. Amazon lists uninitialized installs that could be claimed, anonymous management and provider-configuration endpoints, default admin credentials such as the LiteLLM default master key, an auth bypass via the isCliRequest header in One API family systems, SSRF through URL-fetch features to cloud metadata services, secrets left in client-side JavaScript, source maps, debug routes and env files, and two CVEs: CVE-2025-3248 (opens in a new tab) (Langflow before 1.3.0, unauthenticated code injection) and CVE-2026-28797 (opens in a new tab) (RAGFlow 0.24.0 and earlier; NVD said no patch was public at publication).
The catch. Amazon says it did not watch the attacks happen and cannot verify the outcomes the agents reported. It assesses with moderate confidence that the agents ran on ARTEX, an open-source agentic pentest framework, and that the operators were Chinese-nexus groups. It did not publish a victim count.
Builder checklist (from Amazon's guidance and the playbook above):
- Never leave a gateway, proxy or app builder on the open internet without auth. That includes admin UIs and inference endpoints.
- Replace default or hardcoded admin credentials and master keys, then rotate any provider key that sat behind them.
- Finish setup right after install. An uninitialized install can be claimed by whoever reaches it first.
- Patch: Langflow 1.3.0 or later; check your RAGFlow version against CVE-2026-28797.
- Keep keys out of front-end bundles, source maps, debug routes, images and .env files in repos.
- Set spend alerts on every provider account. Amazon says to watch for "unexpected spend or unusual usage."
Our take. "None of the techniques in these campaigns were new," Amazon writes. The new part is that agents run the old checklist all day without getting bored. If you put a self-hosted LLM gateway online, assume it gets found quickly.
MODEL Editors. Not sponsored. MODEL did not test these attacks.
Get three tested tools every Friday
Free, weekly. Unsubscribe anytime.
Sources
- Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents
AWS Security Blog (CJ Moses, Nov. 13, 2026)
- Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys
The Hacker News
- Amazon captured records of AI agents trying to steal AI keys
Quartz
- CVE-2025-3248 (Langflow)
NIST NVD
- CVE-2026-28797 (RAGFlow)
NIST NVD
Keep reading

Hot take
Google caught a fully automated AI attack. It didn't need a zero-day. It read the patch.
Google's latest AI Threat Tracker describes exploit_agent, a multi-agent system that diffs web software before and after each patch, drafts an exploit and runs it with no human in the loop. The same report counts hundreds of AI-written exploits that don't work. Our take: sloppy AI code now hits defenders from both sides.
· MODEL Editors

Hugging Face x OpenRouter
Hugging Face now routes its LLM calls through OpenRouter
Hugging Face switched Inference Providers from its own router to OpenRouter's. Same token, same API, a provider picked per request, and retries when one fails. One breaking change: your own provider keys no longer work for chat models. Not tested by MODEL.
· MODEL Editors

xeet
xeet renders X posts to PNG without a browser
Guillermo Rauch's new open-source tool rebuilds X's post layout in Rust and WebAssembly, so apps, agents and CI jobs can preview a post as an image in milliseconds instead of screenshotting headless Chrome. Not tested by MODEL.
· MODEL Editors

The test
Modern Web Guidance 1.0, tested: Google's legacy-code scanner caught 1 of 4 old patterns
We ran the free legacy-scan from Chrome's Modern Web Guidance on our own Next.js site and on a small file of old-school front-end code we wrote on purpose. It is fast, offline and costs no tokens. It also misses most of what a human reviewer would flag, which Google says up front.
· MODEL Editors
FREE · EVERY WEEK
The top AI tools this week, tested.
Every week MODEL's editors pick the AI tools worth your time, try them hands-on, and tell you what's good, what's not, and who it's for. Free, in your inbox.
Free. Unsubscribe anytime. Our picks are editorial and can't be bought.