Researchers have discovered a simple mathematical way to “steer” AI models by directly manipulating internal concept vectors – improving performance while revealing hidden risks. Now AI behavior can be controlled more precisely than ever, but also raises concerns about how easily safeguards can be bypassed.