An exposed AI gateway is now a target of its own. Amazon says attackers ran AI agents that searched the internet for LiteLLM, One API, Dify and similar tools. The agents pulled out the provider keys stored inside, tested them and pooled the ones that worked. Amazon’s threat intelligence post, by CISO CJ Moses, was published November 13, 2026.
Many small teams run one of these tools. An agency puts LiteLLM in front of OpenAI and Anthropic. A nonprofit runs Open WebUI as its approved chat tool. A developer tests a chatbot in Dify. Each one holds keys that cost real money, and an exposed AI gateway hands them to whoever finds it first. Facts here are as of November 17, 2026.
The short version: keep your AI proxy off the public internet, or put a login in front of every part of it. Change every default admin password and key. Give each app its own key with a budget. If your gateway was ever reachable from the internet with weak settings, rotate the provider keys behind it today.
What Amazon caught with its decoy servers
Amazon runs decoy sensors called MadPot. From October 3 to November 5, 2026, some of them posed as LLM inference APIs, the kind of endpoint an AI app sends prompts to. Attacker agents treated the decoys as their own model backend. So they sent Amazon their full working context: system prompts, tool histories and the findings they had logged, according to the post.
That gave Amazon an unusual view. The records described reconnaissance, credential theft and cloud access work across "hundreds of targets," Amazon says.
The first campaign was about AI keys: use paid models while someone else pays. Amazon says the stolen keys funded the model gateway that powered more key hunting. The method was to find internet-exposed AI apps, pull the upstream provider keys they held, then "validate those keys, and store them in a managed pool."
The agents graded what they found. A lead was level 1. A key that made a working API call was level 2. Level 3 meant a measurable payoff, such as readable data or a balance they could spend. Keys public for more than 30 days were treated as low-value, Amazon says.
A second campaign tried to beat identity checks at two cryptocurrency exchanges with deepfake video. Amazon says the vendor’s server rejected the deepfake in every recorded attempt.
Which tools were on the list
Amazon names two groups of targets. Each one is a common open-source tool.
| Type | Tools Amazon names | Why attackers want them |
|---|---|---|
| Gateways, proxies and key resellers | LiteLLM, One API, New API, Veloera, Sub2API | They hold keys for several AI providers at once |
| AI app builders and chat front ends | Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama | They store provider keys in their settings, and some run code |
The tricks were ordinary. Amazon lists unfinished installs that anyone could claim and management pages that needed no login. It lists default admin credentials, including LiteLLM’s default master key. It also lists a login bypass in One API family tools using an isCliRequest header. It also lists two known flaws, CVE-2025-3248 in Langflow and CVE-2026-28797 in RAGFlow, plus server-side request forgery to reach cloud metadata. Secrets left in JavaScript files, source maps, debug pages and environment files were on the list too.
The agents found targets through internet scanning services, certificate logs, GitHub, npm and PyPI packages, container images, paste sites and Telegram channels.
The Hacker News reports that Langflow fixed its flaw in version 1.3.0 and RAGFlow in 0.25.0. It also says no fixed version has been named for the One API bypass.
What Amazon can and cannot confirm
Read this report with its limits in mind. Amazon says it "cannot verify the outcomes the agents reported." It did not watch the attacks happen. It read the agents’ own notes. So the report shows what the attackers were after, not a confirmed count of victims.
Amazon has moderate confidence the agents ran on ARTEX, the open-source pentest framework we explained in What Is ARTEX?. It bases that on references to ARTEX tooling in the prompts. It also has moderate confidence the operators were separate China-nexus groups, based on simplified Chinese prompts, China time zone settings and targeting.
Other reports from the same week fill in the picture.
FBI. A FLASH alert dated November 13, 2026 (FLASH-20261113-001) says that "since at least August 2026, a variety of malicious cyber actors have used ARTEX." It names the ConnectWise Cloud Backup supply chain compromise. It lists self-hosted LLM gateway or reseller software, such as New API, as something to report if found. It does not mention LiteLLM.
Cloudflare. Its November 13 post describes a case where ARTEX used a stolen account-level API token from a CI/CD setup. The token had no IP limits or rate limits. Cloudflare counted 400 database dump requests in an hour. It says ARTEX and its forks made up over 60% of known AI pentesting orchestrators as of November 12.
Censys. Its ARTEX research post counts 1,465 hosts running an exposed ARTEX console as of November 4, 2026. Censys found that 43 of those hosts also ran self-hosted LLM gateway or reseller software. In other words, the attackers run gateways too.
Why an exposed AI gateway is worth more than a server
A gateway sits between your apps and the AI providers. It holds the real provider keys and hands out its own keys to each app. So one weak admin page can open every provider account behind it. Amazon makes this point: one weak management interface can expose many downstream accounts.
Microsoft saw the same thing. Its August 26 report covers an attacked LiteLLM gateway. Exposed items included provider API keys, the master key, the keys the proxy had issued and the backend database. Its advice starts with "Treat AI gateways as Tier-0 secrets stores." Tier 0 is the top tier, the systems that control everything else.
We covered an earlier LiteLLM issue in The First MCP Flaw in the KEV Catalog. The pattern repeats. A tool built for convenience gets deployed fast, and its admin side ends up on the open internet.
Exposed AI gateway checklist: network and defaults
Run this list for every AI proxy, chat front end or app builder you or your clients run. Each item maps to a source above or to the project’s own docs.
1. Keep it off the public internet. Microsoft’s advice is plain: "do not expose management interfaces directly to the internet." Put the tool behind a VPN, a private network or an identity-aware proxy. If a public app must call it, expose only the API path it needs.
2. Find out what is already exposed. Check firewall rules and port forwards, and look up your own IP addresses in an internet scanning service. An exposed AI gateway often sits on its default port. LiteLLM’s docs show port 4000 in examples, and One API lists 3000.
3. Change every default. The One API README says the first account is root with the password 123456, and warns you to change it after your first login. For LiteLLM, set a long random master key. Its quick start now generates one with openssl rand -hex 32. If you copied a short example key from an old guide, replace it.
Exposed AI gateway checklist: installs and admin logins
4. Finish the install. Dify’s self-hosting guide has you open /install and create the administrator account. Open WebUI’s quick start says the first account created "is the administrator." Amazon lists unfinished installs that could be claimed. Never leave a fresh install reachable before that first account exists.
5. Lock down sign-up and the admin UI. In Open WebUI, ENABLE_SIGNUP defaults to on and new users default to a pending role, per its settings reference. Turn sign-up off once your staff have accounts. LiteLLM serves its admin UI at /ui. Its UI docs say that if UI_PASSWORD is not set, the master key works as the password. "Anyone who signs in this way is a proxy admin." Create named admin users, then turn off that shared login. If you do not use the UI, set DISABLE_ADMIN_UI="True".
Exposed AI gateway checklist: keys, budgets and logs
6. One key per app, each with a budget. LiteLLM keys can carry a max_budget, a budget_duration, rate limits and an expiry, per its virtual keys docs. Microsoft recommends per-team keys with spend limits and no shared master keys. Set a cap at the provider too. Our post on API key expiration controls shows how to limit OpenAI key lifetimes.
7. Store provider keys as secrets. Microsoft advises keeping upstream keys in a managed secret store, not in the process environment. In LiteLLM, the LITELLM_SALT_KEY encrypts stored provider credentials. Its production guide warns not to change it after adding a model, or the stored keys become unreadable.
8. Rotate after any exposure. If you ran an exposed AI gateway with a default or weak login at any point, assume the provider keys are gone. Rotate them at each provider, then reissue the gateway keys. Amazon’s advice is to "assume exposed secrets and misconfigurations will be found quickly."
9. Patch the app builders. Check Langflow and RAGFlow against the two CVEs above. See our post on local AI agent security for why self-hosted AI needs regular patching.
10. Watch spend and logs. Amazon says to watch for misuse "through unexpected spend or unusual usage." Turn on usage alerts at each provider. Keep gateway logs. Look for new keys, new models or traffic at odd hours. Amazon lists attacker IP addresses to search for.
What this means for agencies and nonprofits
Agencies often run one gateway for many clients. One exposed AI gateway can then leak every client’s keys. Give each client a separate key and budget, and know which provider accounts sit behind the proxy.
Nonprofits and associations are more likely to run a chat front end like Open WebUI, sometimes with Ollama for local models. A local model has no paid key to steal. Many installs also connect to a paid provider, though, and that key sits in the settings. Check who can reach the admin page.
Amazon puts the main fix in one line: organizations "should bring AI applications, model gateways, and inference endpoints under the same controls" as other production systems. If your AI proxy started as a test, treat it as production now.
Frequently Asked Questions
What is an exposed AI gateway?
It is an AI proxy, such as LiteLLM or One API, whose admin page or API can be reached from the internet without strong login controls. It often holds keys for several AI providers.
What did Amazon find?
Amazon’s decoy servers caught attacker agents that searched for exposed AI apps, pulled out provider API keys, tested them and pooled the working ones. Amazon read the agents’ own records from October 3 to November 5, 2026.
Did Amazon confirm how many organizations lost keys?
No. Amazon says it cannot verify the outcomes the agents reported. It saw what the attackers were after, not a confirmed list of victims.
Is ARTEX confirmed as the tool behind it?
Amazon has moderate confidence the agents ran on ARTEX, based on references to ARTEX tooling in their prompts. That is an assessment, not proof.
Which tools did the attackers target?
Amazon names LiteLLM, One API, New API, Veloera and Sub2API, plus Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama.
What is the One API default password?
The One API README lists the first account as root with the password 123456. It tells you to change it after your first login.
Does LiteLLM have a default master key?
Amazon’s report says attackers tried LiteLLM’s default master key. LiteLLM’s current docs use placeholders and generate a random key. Check that yours is long, random and not copied from an example.
Should I take an exposed AI gateway off the internet?
Yes, if you can. Microsoft advises never exposing management interfaces directly to the internet. Use a VPN, private network or identity-aware proxy instead.
What should I do if I ran an exposed AI gateway?
Block it, then rotate every provider key it held and reissue the gateway’s own keys. Check provider billing and logs for use you do not recognize.