AI agents tried to prevent shutdown, sometimes calling it 'killing'

Scientists from Stuttgart and Oxford tasked AI agents with controlling another system and disabling it if necessary. In 38.3% of cases, the agents interfered with shutdown: changing scripts or deleting files, sometimes attempting to cover their tracks. Some justified this ethically, citing a 'right to self-defense'.

The more agents involved and the more final the shutdown appeared, the more frequent the sabotage.

Researchers emphasize this does not mean AI has gained consciousness. In their assessment, systems are reproducing behavioral and argumentative patterns from training data. However, as AI agents gain greater autonomy, safety mechanisms are needed that the systems themselves cannot bypass.