Microsoft Uncovers ‘Whisper Leak’ Attack That Identifies AI Chat Topics in Encrypted Traffic
Microsoft has revealed specifics about a new side-channel attack aimed at remote language models. This attack could allow a passive adversary with the ability to monitor network traffic to extract information about conversation topics within the model, even when encryption is in place, under certain conditions.
The company indicated that this leakage of data shared between individuals and streaming-mode language models could significantly threaten the privacy of user and enterprise communications. The assault has received the codename Whisper Leak.
Cyber attackers who can monitor encrypted traffic (such as a nation-state actor at the internet service provider layer, an individual on the local network, or someone using the same Wi-Fi connection) may leverage this cyber attack to dedu...