LLM Agent Honeypot: Monitoring AI Hacking Agents in the Wild

A preprint on an LLM agent honeypot for monitoring AI hacking agents in the wild. The companion cites its result that the honeypot picked out eight potential AI agents among more than eight million hacking attempts, by planting prompt injections that only an agent would obey. It is an automatic defender, an enemy of parasites in the food web, and a field instrument for detecting agents.

Cited in

the companion paper (Companion §4)

Used for

Food web (enemies of parasites); field detection

Strand

Agents in the wild: incidents & field evidence · Misalignment reports & field detections

Cite

@misc{reworr2024llm,
  title = {LLM Agent Honeypot: Monitoring AI Hacking Agents in the Wild},
  author = {Reworr and Dmitrii Volkov},
  year = {2024},
  url = {https://arxiv.org/abs/2410.13919},
}