Attackers' AI Agents Are Stealing Keys From LiteLLM and Ollama. 3 Fixes for MSPs.
The video uses a digital avatar of Erick Grau, CEO of Chibitek.
If you run an AI gateway for your clients, it is now a key vault. And attackers' AI agents are shopping for it.
A lot of us have stood up LiteLLM, Ollama, Open WebUI or something like it lately, often fast, often as a pilot. This post is about what happens when one of those pilots is left reachable from the internet. I re-checked everything below against Amazon's report, the FBI's alert on the same framework, and the LiteLLM and Ollama documentation on November 18, 2026.
What Amazon caught
On November 13, 2026, Amazon Threat Intelligence published what its MadPot decoy sensors captured from October 3 to November 5, 2026. MadPot is Amazon's network of decoy sensors. For this work, some of them posed as popular AI model APIs.
The visitors were AI agents. Amazon assesses, with moderate confidence, that they ran on ARTEX, an open-source agentic AI penetration testing framework. Because these frameworks send their working history along with each request to an AI model, the requests that hit the decoys carried the agents' own records of their campaigns.

The main campaign followed one method, which Amazon spells out:
Find internet-exposed AI applications.
Pull out the upstream provider keys they hold.
Check that the keys work.
Store them in a shared pool.
The targets included AI gateways and proxies such as LiteLLM, and AI app builders such as Dify, Open WebUI and Ollama, along with several others Amazon names.
Then comes the part that should stick with you. Amazon describes the operation as self-sustaining: the stolen keys paid for the AI that powered the hunt for more keys. Your client's AI bill becomes the attacker's fuel.
Why gateways are the prize
Amazon says the agents focused on applications that aggregate provider keys, "because one weak management interface can expose many downstream accounts."
Read that again as an MSP. A gateway is useful precisely because it holds keys for several providers and serves several teams or clients in one place. That also makes it one door to many accounts. The design choice that saves you time is the same one that makes a weak admin page expensive.
None of the techniques were new
This is the line from Amazon I would put on the wall: "None of the techniques in these campaigns were new. What AI agents change is the speed and persistence of the attacker."

The agents carried a playbook that mapped each product to known ways in. The entries an MSP can act on this week:
Default administrator credentials, such as the LiteLLM default master key.
Uninitialized installs that could be claimed. Someone starts a setup, walks away, and the first visitor gets to finish it.
Management and provider settings endpoints with no login.
Keys left in code and public files, including client-side JavaScript, source maps, debug routes and environment files.
Two product details are worth knowing, straight from the vendors' own docs:
LiteLLM: out of the box, the Admin UI accepts the username admin and, if no separate UI password is set, the master key itself as the password. LiteLLM calls this a bootstrap mechanism only and documents a setting, disable_env_credential_login, to turn it off once real admin accounts or SSO are in place.
Ollama: it binds to 127.0.0.1 port 11434 by default, and its local API does not require authentication. That is fine on a laptop. It stops being fine when someone changes the bind address so another machine can reach it, or publishes the port from a container on a server with a public address.
One honest caveat, and Amazon makes it too. The decoys captured the agents' records, not the attacks themselves. Amazon "cannot verify the outcomes the agents reported," and it notified AWS customers who may have been affected. So I am not going to give you victim counts. But the playbook is real, and every item on it is something we can check.
This is not a one-off
On the same day, November 13, the FBI published a FLASH alert on ARTEX. The FBI says a variety of attackers have used it since at least August 2026 against organizations across many sectors, and that the original developer removed the GitHub repository in October, but forks, archived copies and a commercial version keep it available. Among the campaigns the FBI cites is a supply chain compromise affecting ConnectWise Cloud Backup customers. If you needed a reason to take agentic attack tools seriously as an MSP, that one is close to home.
Two of the FBI's recommendations line up with the fixes below: restrict administration interfaces and other sensitive services to trusted networks or dedicated admin paths, and rotate exposed credentials promptly.
Three fixes for your AI gateway
1. Change every default admin credential, and finish or tear down any install you walked away from. On LiteLLM, that means real per-person admin accounts or SSO, then turning off the environment credential login so the master key is no longer an Admin UI password. Then go hunting for the half-built installs: the test box from the pilot, the container someone spun up for a demo. If it is not finished and owned, tear it down.
2. Get the admin page and the inference endpoint off the open internet, behind a login. Amazon's guidance is to bring AI applications, model gateways and inference endpoints under the same controls as your other production systems. For Ollama, leave it on loopback unless you have a reason not to, and if you need remote access, put an authenticated layer in front of it. Then test from outside your network, not from the box itself.
3. Treat a leaked key as found the same day. Amazon says to assume exposed secrets and misconfigurations will be found quickly, and the FBI says to rotate exposed credentials promptly. My rule for our shop is simple: if a provider key shows up somewhere it should not be, rotate it that day. Amazon also says to watch for misuse of AI services through unexpected spend or unusual usage, so set spend alerts on every provider account. You want a stolen key to show up on your bill before it shows up in the news.
Key Points
Amazon's decoys captured AI agents, tied with moderate confidence to the open-source ARTEX framework, hunting internet-exposed AI apps such as LiteLLM, Dify, Open WebUI and Ollama from October 3 to November 5, 2026.
The agents pulled out provider keys, checked them, pooled them, and the stolen keys paid for more hunting.
Gateways were targeted because one weak management interface can expose many downstream accounts.
None of the techniques were new: default admin credentials, unfinished installs, endpoints with no login, keys left in code.
Amazon cannot verify the outcomes the agents reported, so there are no victim counts here. The playbook is still worth checking against.
Fixes: change default admin credentials and clean up abandoned installs, keep admin pages and inference endpoints behind a login and off the open internet, and rotate leaked keys fast with spend alerts on.
Frequently Asked Questions
We only run Ollama on one internal server. Are we exposed? By default Ollama binds to 127.0.0.1, so only the same machine can reach it. Check whether anyone changed the bind address or published the port from a container, and test from outside your network. If it is reachable, remember its local API does not require authentication.
Is the LiteLLM master key really the admin password? By default, yes, if no separate UI password is set: the Admin UI accepts the username admin with the master key. LiteLLM documents how to move to per-user accounts or SSO and then disable that login path.
Did these agents actually steal keys from real companies? Amazon says it cannot verify the outcomes the agents reported. What it could see was the method and the targets, and it notified AWS customers who may have been affected.
What is ARTEX? Amazon and the FBI both describe it as an open-source, AI-driven penetration testing framework that attackers have turned into an attack tool. The FBI says it has been used since at least August 2026.
What should I do first? Inventory. Find every AI gateway, AI app builder and model server you or your clients run, including pilots, and check which ones answer from the internet.
Let's Compare Notes
At Chibitek, we are an MSP talking to MSPs. AI gateways are new enough that most of us are still writing the playbook for running them. Three questions to answer about your own environment: do you know every AI gateway and model server you run, can any of them be reached from the internet without a login, and would you notice a provider key being used by someone else before the invoice arrived? If you want a second set of eyes, we are happy to compare notes. Contact us at hello@chibitek.com.
We make IT effortless so you can create, disrupt, and grow!
Follow Chibitek on YouTube at @chibitekmsp, and follow Chibitek on LinkedIn.








Comments