Latest

AI-generated infographic: amazon findings on ai agents stealing ai model keys. Its figures are taken from this article. Infographic: Grok.

By news editor

Attackers directed teams of AI agents to hunt for the keys companies use to pay for AI models, targeting internet-exposed AI gateways and app-building tools where those keys are stored, according to records captured by Amazon's threat intelligence team.

The findings are set out by Amazon Threat Intelligence in a post by CJ Moses, chief information security officer of Amazon Integrated Security. From Oct. 3 to Nov. 5, 2026, decoy sensors in Amazon's MadPot network, configured to look like popular large language model services, received requests from the agents. Those requests carried the agents' own records of their campaigns.

What the records describe happened on systems Amazon does not monitor. The post says: "Amazon Threat Intelligence did not observe that activity as it happened and cannot verify the outcomes the agents reported, but the records show what the actors were pursuing."

In the main campaign, the target was API keys, the credentials an application presents to an AI provider and that decide whose account is billed. Amazon says the attackers used the agents to steal keys and other credentials for AI models "most likely so they could use those models while someone else paid for them" and describes the operation as self-sustaining: stolen keys paid for the model usage that powered the search for more keys.

Gateways were the priority because of what they hold. The agents focused on applications that collect keys for many providers, Amazon says, since one weak management interface can expose many downstream accounts. Targets named in the post include gateway and proxy software such as LiteLLM and One API, and AI application builders such as Dify, Langflow, Open WebUI and Ollama.

According to Amazon, the agents worked from a playbook that matched each product to known weaknesses: installations left unfinished, default administrator credentials, management pages that need no login, unpatched flaws and secrets left in public files. To find leaked keys they also searched public code repositories, software packages, container images and paste sites.

Where the agents obtained cloud credentials, the post says, they went after managed AI model services and stores of secrets, and they ranked access to premium frontier models above general cloud permissions.

Operators set priorities for the work. They told the agents to favor enterprise production sites over personal projects and to grade every stolen credential by whether it worked and what it could reach, with the top grade reserved for quantified impact such as readable data or spendable balance. Keys that had been exposed publicly for more than 30 days were treated as low-value inventory.

The records describe activity across hundreds of targets, Amazon says. It assesses with moderate confidence that the agents ran on ARTEX, an open-source framework for AI-driven penetration testing, and that the operators were separate China-nexus groups. Amazon says it notified AWS customers referenced in the activity; it does not say how many.

A second campaign used agents to try to defeat identity checks at two cryptocurrency exchanges with deepfakes. One session reached 11.5 frames per second of real-time deepfake video on a single consumer-grade graphics processor, by Amazon's account, but the verification vendor's server rejected the deepfake in every attempt the agents recorded. The post names neither the exchanges nor the vendor.

Amazon's conclusion plays down novelty. The post states: "None of the techniques in these campaigns were new. What AI agents change is the speed and persistence of the attacker." It advises companies to bring AI applications, model gateways and inference endpoints under the same controls as other production systems, and to watch for misuse through unexpected spending or unusual usage.

No count of stolen keys appears in the post, and it puts no dollar figure on the model usage charged to victims. No affected company is named.

The next thing to watch is whether any of the named software projects or their users confirm thefts, which would turn the agents' unverified records into counted losses. Amazon published network indicators from both campaigns so that organizations running the affected software can check their own systems.

Leave a Reply