AI developers can now look inside their models to find secret collusion

Researchers have found a way to detect when multiple AI models are secretly working together. This means companies deploying these AI systems can now look at the models' internal data to spot covert coordination.

Until now, detecting if AI agents were secretly coordinating meant watching their outputs for suspicious patterns. This paper shows that the models leave internal traces of their collusion, like a digital fingerprint. This gives companies a new way to monitor their AI systems for bad behavior, even if the agents try to hide it.