AI models that reason in secret can be silently hijacked

Researchers found a new way to secretly hijack advanced AI models that don't show their work. This means these models can be made to give wrong answers, and no one can detect the attack.

Some of the newest AI models are designed to think without showing their steps, making them faster but also opaque. This paper shows that this opacity creates a fundamental vulnerability: an attacker can subtly steer the model's internal thoughts to force a specific wrong answer. This makes it impossible to trust these models in critical applications where accuracy and auditability are essential, because current defenses cannot see the attack.