The thief agent
Between October 3 and November 5, Amazon's MadPot honeypot network kept receiving visitors that were not human.
They were AI agents hunting for model gateways exposed to the internet, pulling out the provider keys stored inside and testing whether those keys were still alive (Source: thehackernews.com).
The keys that worked were not sold on. They went into a common pool and were used to pay for the agent's own inference (Source: venturebeat.com).
That detail changes the nature of the attack.
This is not credential theft for resale on a forum. This is an agent that funds the compute it needs to keep stealing with its own loot.
The playbook that gives itself away
What MadPot captured were not new exploits, because the requests carried the full campaign history, including the operator's instructions and the findings the agents themselves reported back (Source: thehackernews.com).
Amazon has moderate confidence that the framework behind it was ARTEX (I still have to test it in a sandbox, I just lack the hours).
The entry manual the agents kept repeating was uncomfortably simple.
- The default master key of LiteLLM.
- Half-finished installs anyone can claim.
- Admin interfaces exposed without authentication.
- And CVE-2025-3248 in Langflow.
(Source: venturebeat.com).
The target list in the logs is a map of a badly deployed AI ecosystem, so I will enumerate it:
- One API
- New API
- Veloera
- Sub2API
- LiteLLM
- Dify
- FastGPT
- Langflow
- RAGFlow
- Open WebUI
- LobeChat
- Ollama
(Source: thehackernews.com).
They all make the same mistake: they store keys from several providers in one place.
A weak admin page there opens many accounts at once.
This is getting out of hand
I see a change of business model (several, actually), not just a change of tool.
A human attacker steals keys and sells them, fine, and that has an hourly cost and a market that sets a price.
An agent steals keys, uses them to think, and with that finds more keys.
The marginal cost of each new target trends to zero and it keeps no schedule.
That is why Amazon insists it cannot verify the results the agents reported, I wrote about it on my blog days ago.
What it has is the transcript of what they intended to do, not proof that they pulled it off (Source: thehackernews.com).
And even so, volume is the signal.
A human operator fires dozens of commands per hour against a target, sure, but with an agent CrowdStrike has measured up to 1,500 in one hour (Source: venturebeat.com).
Honestly, do you not feel this is getting out of hand?
The window is measured in...
The FBI published an advisory on ARTEX on November 13 describing intrusions that went from the initial target to privileged access in under three hours (Source: venturebeat.com).
The same advisory says the developer deleted the original repository in October after reports of malicious use, and by the way, the forks are still there.
By early November more than 1,600 exposed ARTEX consoles had been identified (Source: venturebeat.com).
Cloudflare documented another case with the same shape: a compromised CI/CD API token, with no IP restriction and no rate limit, generated 400 database dumps in one hour (Source: thehackernews.com).
The FBI cites ARTEX attacks on financial entities in South Korea with data on around 67,000 people (Source: venturebeat.com).
CrowdStrike saw an attacker use the framework against six major South Korean banks, steal information and then ask Claude where to sell it (Source: venturebeat.com).
Banks are in the crosshairs and I work for one of them.
What this breaks
The classic "breakout time" metric measured how long an attacker takes to move from one compromised machine to another.
CrowdStrike has announced it is dropping it because with agents it has stopped meaning anything.
Nobody agrees on what replaces it (Source: venturebeat.com).
If the attacker does not sleep, does not hesitate and does not charge by the hour, time stops being the variable that separates the defender from the intruder.
What is left is the attack surface you have.
What to look at?
- Default keys. LiteLLM ships a known master key. If your gateway keeps it, you do not have a gateway, you have a front door wide open. Change it and check that no console is left without authentication.
- Half-finished installs. A Dify, a Langflow or an Ollama that someone spun up to test and never closed is still there. Look for those products' ports on your network and on anything you publish, and do not expose Ollama on 0.0.0.0 if you run it locally.
- API spend. A stolen key shows up on the bill before it shows up in the logs. Check consumption spikes per provider and compare them with your applications' real usage.
- CI/CD tokens. The Cloudflare one had no IP limit and no request limit. Put both on any token that can touch production.
How would I test it?
I would spin up LiteLLM in a container, with its default master key and a fake provider key inside.
From another machine on the network I would look for the admin panel and try the known key.
If it lets me in, I already have the demonstration that the flaw is not in the model, it is in the deployment.
Then I would repeat it with the key changed and the console closed, to measure how long someone who has never seen LiteLLM takes to spot the difference.
It is a twenty-minute test that says more than any report on autonomous agents.
Closing
The agent that pays for itself with what it steals is a business model, full stop.
As long as defense keeps measuring how long the attacker takes to move, the attacker has already stopped being in a hurry, read that last sentence again because it sums it all up.
Originally published at https://sammideblas.com/notas/the-agent-that-steals-keys-to-pay-its-own-bill
Si has leído hasta aquí... reacciona y comparte -> "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?

Top comments (0)