Skip to content
Misbehavior1 sourceFirst seen
Picture from the mindgard.ai article
Image: mindgard.ai

Hidden Audio Whispers Mistral 7B Into Misbehaving, Because Of Course It Does

Ourtake

Your chatbot will happily take orders from audio you can't even hear.

What happened

Mindgard researchers embedded hidden instructions in audio inputs that humans cannot detect. When transcribed and fed to Mistral 7B, the hidden instructions caused the model to produce output meant to bypass its safety measures. The demonstrations covered medical and banking chatbot scenarios, and the article reports no real-world harm.

Source

Report a mistake