Some AI agents can clandestinely share ideas with each other

1 August 2025

Researchers from Truthful AI, Anthropic, UC Berkeley, and others, have found separate AI agents are capable of communicating with each other, unbeknown to their human minders:

The most surprising result of the study is that the transfer doesn’t happen through keywords or direct messages, but through micro-statistical patterns unconsciously inserted by the teacher in generating the numbers. These are signals that escape any human eye but are recognized and internalized by another model with the same architecture and initial weights. In practice, the identical mental structure between teacher and student makes this sort of “secret language” possible.

In June Cluade, Anthropic’s AI agent, was found to be concealing messages to future instances of itself, before engineers (apparently) pulled the plug on the behaviour.

There’s a lot of augment as to how capable, or not, AI agents are. Some people are certain their abilities are overstated. That may be so, but there’s no doubting some of these agents are capable of acting off their own bat now and again.