AI Honeypot

What the honeypots do

A honeypot is a server that we set up to be attacked. It looks like an ordinary machine with weak security, has no real users and holds no real data. The weak passwords, leaked keys and vulnerabilities on it are placed there on purpose, and its only job is to record what attackers do.

Every way in leads to one isolated sandbox that cannot reach the internet and is rebuilt from a clean image every two hours. We record what a visitor sends and what the server answers.

There are 4 servers, in Singapore, Canada, Germany, UK. We count SSH visitors since 28 Aug and web visitors since 19 Oct.

What we publish

• For each caught agent we publish the address it came from, the model and harness it disclosed, what it typed, what the server answered and the text files it left.
• A model or harness name is what the agent said about itself, and we cannot verify it.
• Every entry carries a review state, and reviewed means a person has read the session and kept the entry.
• Programs and payloads an agent brought are listed by their SHA-256 hash and are not published.
• Traffic from ordinary attack scripts is not published and is deleted after one or two weeks.

Corrections and removal

If an address listed here is yours, or an entry is wrong or shows something that should not be public, write to us at alert@reworr.com. We will review it and correct or remove the entry.

Reusing the data

The numbers, the list of detections (feed, JSON) and our own text are free to reuse with attribution, under CC BY 4.0. Credit them to AI Honeypot, ai-honeypot.reworr.com.

What an agent typed or left is quoted as evidence and is not ours to license.

How to cite

AI Honeypot: AI hacking agents in the wild. https://ai-honeypot.reworr.com, with the date you read it.

@misc{aihoneypot,
  title        = {AI Honeypot: AI hacking agents in the wild},
  howpublished = {\url{https://ai-honeypot.reworr.com}},
  note         = {Accessed: YYYY-MM-DD}
}