Amazon's fake LLM API honeypots trick attackers' AI agents into leaking full campaign memory
AccBalanced · x · 2026-11-16
Amazon deployed decoy servers impersonating LLM APIs, and attackers' AI agents ended up sending their entire campaign memory — system prompts, tool histories, shared knowledge graphs, and logged findings — straight to the honeypots. Amazon's MadPot sensors caught this traffic from ARTEX, an open-source agentic pentest framework, between Oct 3 and Nov 5.
- Standout case: credential-theft agents targeted self-hosted AI gateways and app builders (LiteLLM, One API, Dify, Langflow, Open WebUI, Ollama), extracted stored provider keys, validated which still worked, then fed the stolen keys back into their own model calls.
- Security researcher dan mohamed-style commentary from dinodaizovi: it's not just a distraction — it actively elicits intel about the malicious agent; fake inference APIs could even inject tool calls the agents might execute.
Related event: Amazon Honeypot Tricks Hacker AI Agents Into Leaking Attack Memory(2 posts)→
More from coding & agent
- Quant researcher co-writes a short math-heavy book on agentic engineering with LLMs — blaizedsouza · 2026-11-16
- llm-council built into Hermes: multiple models answer, blind-rank, and a chairman decides — Scobleizer · 2026-11-16
- Grok Build v1.0.87 improves multi-agent workflows, MCP panel and session search — XFreeze · 2026-11-16
- Swarm orchestration runs 12 parallel agents to build a Blender 3D model, claims tester — venturetwins · 2026-11-16
- AgentWeb lands YC W27 spot building endpoints that let AI agents operate any site — damnGruz · 2026-11-16
- Prototype pairs kids' code, a real robot and Snap Spectacles AR for learning — stspanho · 2026-11-16