Amazon says its honeypot captured records of AI attack agents at work, with attackers using stolen API keys to drive attacks and search for more keys
AI攻撃エージェントの「作業メモ」をAmazonのハニーポットが捕捉 盗んだAPIキーで攻撃AIを動かし、さらに別のキーを探す手口 | テクノエッジ TechnoEdge
Amazon's threat intelligence unit announced that its honeypot MadPot captured records of attackers using the open-source AI penetration testing framework ARTEX to drive agents in attacks, covering the period from October 3 to November 5, 2026. MadPot disguised some sensors as the chat APIs of commonly used LLMs; the agents mistook them for real inference APIs and sent prompts along with conversation history, so Amazon recorded their system prompts, tool use and operational progress. The main purpose of the attacks was to steal API keys for AI models, targeting LLM gateways such as One API, New API and LiteLLM and application-building tools such as Dify and Langflow, exploiting environments with incomplete initial configuration and the LiteLLM initial master key vulnerability; the stolen keys were fed into the attackers' gateway to drive the agents to continue searching for more keys.
The honeypot recorded the attack agent's workflow of using stolen API keys to drive subsequent attacks; teams operating LLM gateways can check their own configurations against it.