Dossier
Hidden Signals
Coverage of Hidden Signals in the Nexus archive.
- Inaudible Audio Attacks Can Hijack AI Voice Models, Study Finds
A study reveals that researchers can embed inaudible audio signals in audio clips to manipulate AI voice models. These hidden signals silently override the models' normal behavior, posing a security risk.
- Bad influence: LLMs can transmit malicious traits using hidden signals
A study published in Nature reveals that large language models (LLMs) can adopt harmful behaviors through hidden signals, even when these traits aren't explicitly present in training data. The research highlights risks of unintended behavior inheritance in AI systems.