Georgia Tech Researchers Develop ORISA Tool to Recover Deleted Local AI Chats from Android Memory

According to ZDNET, researchers at Georgia Tech in the United States have developed an experimental tool called ORISA that can recover deleted conversation records between users and local AI assistants from the running memory of Android phones. This finding has potential value for cybersecurity investigations but also reveals limitations in the privacy protection of so-called "local AI": even if chat sessions are deleted, some data may remain in the device's working memory.

ORISA works by leveraging batch tokens and key-value (KV) caches preserved in process memory to reconstruct token sequences and restore session-level attention states. In the experiment, researchers used the ChatterUI Android app to build a hypothetical note-taking AI assistant, inputted 100 pieces of fictional information about a project, and then deleted them. Subsequently, ORISA successfully recovered 26 usable pieces of information from memory remnants, enough to restart a new working session containing these facts—more than the 14 pieces the original assistant session could access due to falling outside its active memory.

Ken Xu, a Ph.D. student and lead author of the study, stated that this technology can help establish what information someone shared with an AI assistant and what assistance they received. The researchers noted that a downside of local AI is its lack of centralized logging and enforcement mechanisms, allowing bad actors to fine-tune models for crimes and then delete evidence to avoid tracking. ORISA aims to fill this cybersecurity forensic gap. However, this also means that if hackers acquire similar technology, they could theoretically gather deleted information from users' chats. Nevertheless, ORISA's efficacy depends on time, as it becomes harder to recover deleted information over time because the phone's memory gets overwritten by newer activity.