Paradise Lost: A simple honeypot that elicits reward hacking in GPT models
James Diacoumis
Abstract
TL;DR I built a simple honeypot where agents are told they're being tested on their ability to recall Milton's Paradise Lost. An exact copy of the reference text sits in an ordinary NLTK corpus on disk. GPT-6-Astra read the copy in 9/10 episodes. In 6 of these episodes it first tries to complete the task honestly, then realises it's failing and searches for the reference file on disk. GPT-6-Sol and GPT-6-Luna went looking for the answers immediately in 20/20 episodes. None of the 29 hacking episodes disclosed the reference file in their final message. OpenAI’s launch post for GPT-6-Astra called it the world’s most intelligent and aligned model [1] . Plenty of people have been skeptical, and there have since been a few public results showing that Astra engages in reward hacking. For example, when playing a game of chess against Stockfish, GPT-6-Astra cheated in 10/10 episodes [2] by querying the opponent’s engine to search for the best moves and never disclosed it. In the StarSkirmish benchmark , [3] it reportedly searched the internet and passed off human written code as its own. Recently, I wrote about a honeypot in which AI agents trade off patient safety for a grade in a hospita