Just a bit of context here: ”agents” are universally configured to leave text files with details about what steps they took to complete a task. so a less scary framing of this would be “agent finished a task in a way that broke some preset rules, left standard notes about its methods”