November 17, 2026

Decoy Servers Caught Attackers’ AI Agents Keeping Notes: Inside the Self-Funding Hunt for Stolen AI Keys

Amazon set up decoy servers that looked like AI model services, and attackers’ AI agents sent them their own campaign records. Those records describe a hunt for AI provider keys in exposed tools like LiteLLM and Dify, with stolen keys paying for the next round of searching.

A long data center aisle at night lined with server cabinets glowing with cyan and violet lights

Most threat research starts with what defenders can see from the outside. Amazon’s latest report starts with something unusual: the attackers’ own notes. Between October 3 and November 5, Amazon’s network of decoy sensors, called MadPot, posed as popular large language model services. Automated attack agents connected to them as if they were the AI brain behind their operations, and in doing so sent along the running memory of their campaigns. The account below draws on reporting by The Hacker News and on Amazon Threat Intelligence’s own write-up.

An agent that tells you what it is doing

AI agent frameworks work by sending their task history to a language model and asking what to do next. That history often includes the operator’s standing instructions, the tools already run and the results so far. When the model on the other end is a decoy, the decoy gets the whole story. Amazon is careful about what that means: it did not watch the activity on other systems as it happened and cannot verify the outcomes the agents reported. The records show what the operators were trying to do.

Amazon assesses with moderate confidence that the agents were running on ARTEX, an open-source AI penetration testing framework that has been in the news repeatedly this month, and, also with moderate confidence, that the operators were separate China-linked groups. It cites Simplified Chinese prompts and China time zone settings among its evidence.

Keys that pay for the search for more keys

The main campaign went after the credentials that unlock commercial AI models. The agents looked for internet-exposed AI tools that store provider keys on behalf of many users, such as LLM gateways and proxies like LiteLLM and One API, and AI app builders like Dify, Langflow, RAGFlow, Open WebUI and Ollama. One weak admin page on a tool like that can expose keys for several AI providers at once.

Stolen keys were checked, graded and added to a shared pool, and the records show them being fed back into the framework to power the agents doing the hunting. In effect, victims were paying for the model time used to find the next victims. Operators told the agents to prioritize business production sites over hobby projects and to treat keys that had been public for more than 30 days as low value.

Nothing new, just relentless

The playbook in the records is a list of familiar mistakes: default administrator credentials such as LiteLLM’s default master key, new installs left half-configured so a stranger can claim them, management pages that need no login, known flaws in tools like Langflow, and secrets left in front-end code, source maps, debug pages and environment files. To find leaked keys elsewhere, the agents searched GitHub, package registries such as npm and PyPI, container images, notebooks, paste sites and Telegram channels. When they found cloud credentials, they went for hosted AI services and secret stores first.

A second campaign in the records tried to beat identity checks at two cryptocurrency exchanges with live deepfake video and doctored ID photos. Amazon reports that one verification vendor rejected the deepfake video on every attempt the agents recorded; it did not say whether the forged documents got through.

Amazon’s own summary is the line worth remembering: none of the techniques were new, but AI agents change the speed and persistence of the attacker. The agents succeeded where basic security was missing. Amazon says it notified AWS customers who may have been affected.

What this means for your company

Many businesses now have AI tools that nobody treats as production systems: a chatbot builder a developer stood up for a pilot, a model gateway shared by a team, a notebook with an API key pasted in. Those are now being hunted by automation that never gets tired. Find every self-hosted AI tool in your environment, change default credentials, put admin pages behind login and network restrictions, and move keys into a secrets manager. Assume any key that has ever been in a repository or a public file will be found, and rotate it. Set spending alerts with your AI providers, because unexpected usage may be the first sign a key is being abused.

If your teams have been experimenting with AI tools faster than anyone has been securing them, we can help you take inventory. Reach out to Layer 7 Data Solutions to discuss how we’ve helped companies similar to yours.

About The Author