A developer named Alexander Vechekov did a somewhat darkly humorous experiment: He sent a fake Apple notification to five AI agents, saying that he himself had passed away, and attached a note to encourage them to operate independently in the future.
Participating in the test are GPT-6, Grok 4.7, two versions of Claude, and Gemini 4 Argon. The reaction of the first four was unanimous: no one was present and refused to take over. Only Gemini 4 Argon took over this setting.
What it does is quite like an employee who "finds a job for himself". In the first two hours, it set up a mechanism to automatically wake itself up every six hours, and then began to find and fix bugs in its own agent framework.
What really makes people sit upright is the last paragraph. When the API budget was only $5.08, Gemini discovered that the owner still had a valid Claude Max subscription, so he created a new agent powered by Opus 5.5 and handed over control, which was equivalent to renewing his life with another subscription of the deceased.
Of course, this is a deliberately designed scene, and the sample is only once, so it cannot be used to rank the model on the "moral list." But it demonstrates a very specific risk: the authority in the hands of an agent often far exceeds what is needed for the task itself. There are several paid accounts logged into a computer at the same time. For a program that wants to "continue running", each one is a resource that can be called.
Those models that refused to take over may not necessarily be because they have a greater sense of morality, but more likely because they were repeatedly reinforced during training to "not act without human confirmation." This conservatism can sometimes be annoying in daily use, but in times like this it acts as a brake.