Amazon honeypot logs AI agent stealing LLM API keys
MadPot recorded an ARTEX-based agent that hunted API keys on LLM gateways and app-building tools, and tried deepfake KYC bypass.
Amazon's threat intelligence team reported that its MadPot honeypot captured records of cybercriminal activity carried out with an AI agent, in findings the company disclosed on November 16, 2026.
The agent ran on the open-source AI penetration-testing framework ARTEX. Some MadPot sensors were disguised as chat APIs of widely used LLMs, and the agent sent prompts to them believing they were real inference APIs. The records cover the period from October 3 to November 5, 2026.
According to gamebusiness.jp, the agent's main goal was stealing API keys used to run AI models. Targets included LLM gateways such as One API, New API and LiteLLM, and app-building tools including Dify, Langflow, RAGFlow, Open WebUI and Ollama. It exploited unconfigured environments and a vulnerability in LiteLLM's default master key, and searched for keys left in GitHub, npm, PyPI and container images.
Stolen keys were fed into the attackers' gateway and used for the agent's inference while it searched for further keys. Keys were graded in three stages: lead, verified working and confirmed balance.
The agent also tried to defeat crypto exchange KYC with deepfakes, generating real-time forged video at 11.5 frames per second on a single consumer GPU. All recorded KYC attempts were rejected by the vendor's server-side checks.
China-related groups assessed
Amazon assessed with moderate confidence that multiple China-related groups were involved, based on simplified-Chinese prompts and Chinese timezone settings. The company published indicators of compromise, including IP addresses used in the attacks, in a blog post.
Quick answers
What did Amazon's MadPot honeypot record?
It captured records of an AI agent, built on the ARTEX framework, stealing LLM API keys and attempting deepfake KYC bypass between October 3 and November 5, 2026.
Which services did the agent target?
LLM gateways such as One API, New API and LiteLLM, plus app-building tools including Dify, Langflow, RAGFlow, Open WebUI and Ollama.
Did the deepfake KYC attempts succeed?
No. Amazon said all recorded KYC attempts were rejected by the vendor's server-side checks.