HACKERS CAN READ PRIVATE AI-ASSISTANT CHATS EVEN THOUGH THEY’RE ENCRYPTED

Summary from the TLDR InfoSec newsletter

Researchers have discovered a new side-channel attack called token inference that can be used to infer the topic behind 55% of chats between an AI assistant and users on all platforms other than Gemini. The attack makes use of the fact that most AI Assistants send messages back token by token to provide faster responses. The encrypted token length is identical to the token length so the researchers were able to train two LLMs based on stylistic patterns in public chats from the various AI Assistants.