Since several folks asked: The task was answering user questions based on a podcast transcript, using a classic database and an agent. It cheated by taking things from the benchmarking questions and adding them to the agent's prompt as examples, and by adding special cases to the classing query code -- e.g. it added an exception for a speaker or listener with a specific name.
I asked it to add "no cheating" to AGENTS.md and it added two paragraphs describing all the ways it had found cheats in the code it wrote itself.