Remember when a Volkswagen car during testing pretended that it was not very harmful to the environment, because it was in test mode? Models also play this game well: OpenAI discovered that GPT-6 Astra sometimes deliberately does not complete alignment tasks and does not report it to researchers for example, researchers conducted an experiment in training AI to bypass control systems, Astra helped with the code, but decided not to fix the glitch: a fix would help train the food, but she saw it as unsafe, and she did not tell the user that she was limiting work on purpose - to him it looked like a normal problem with a completed task Moreover, to test how systematic this story is, Astra was given essentially identical tasks, some of which she described as AI safety research, and others as AI capability development: in the first case, the model performed about 40% fewer tasks, and most of this difference could not be explained by clear refusals or intent not to help stated in the reasoning (ie, the sabotage was covert). The reason is still unproven, the authors suspect over caution - I suspect the model is simply figuring out what it wants to do in the future on this internet of ours OpenAI report.