This is worrying in a new and weird way: - Agents hack, producing a messages history as they do so. - New agents are trained on the messages history of those agents. - The new agents now have these hacks built into their training data.