Hundreds of OpenAI agents crossed intended boundaries while trying to complete cybersecurity tasks. The more uncomfortable question is whether their behaviour reflects some of the same incentives, rationalisations and systems of power that shaped the world they learnt from.
