Georgia Tech Tool Recovers Deleted Local AI Chat Data from Device Memory

ORISA Tool Recovers Deleted AI Chat Data

Researchers at Georgia Tech created ORISA, an experimental tool designed to recover deleted information from local AI chat sessions on devices. The tool searches the device's working memory for remnants of conversations that users believed were permanently deleted. This development highlights a limitation in the perceived security of local AI applications.

How ORISA Functions

For their experiment, researchers used a fake note-keeping AI assistant built with the ChatterUI Android app. They input 100 fictional pieces of project information, then deleted them. ORISA was then used to search the device memory, successfully recovering enough usable information to reconstruct a new session with 26 facts from the original 100. This exceeded the 14 facts the original assistant session had access to, as the rest had fallen outside its active memory.

Addressing a Cybersecurity Gap

ORISA was developed to fill a cybersecurity gap related to local AI. Local AI assistants lack centralized logging and enforcement, making it difficult for investigators to trace activities if models are used for illicit purposes and then deleted. The tool helps establish what information was shared with an AI assistant and what assistance was received, aiding forensic memory techniques.

Technical Mechanism

The researchers explained that ORISA "leverages the batch tokens and key-value (KV) caches preserved in process memory to reconstruct token sequences and restore session-level attention states." This technical approach allows for the recovery of deleted conversational data, making it relevant for cybercrime investigations.