Georgia Tech’s ORISA tool extracts deleted AI logs from Android memory

Deleting a conversation from an artificial intelligence assistant running on an Android mobile device does not permanently remove every trace of the data from the hardware’s working memory, according to research presented at the 2026 IEEE Symposium on Security and Privacy.

Key Clinical Takeaways:
• Researchers at Georgia Tech developed a recovery tool named ORISA that extracts deleted AI conversation logs from Android device memory.
• The testing phase evaluated 10 distinct Android applications and 3 separate AI models across 60 configurations.
• Data persistence occurs because local AI assistants keep active working memory segments even after users trigger deletion commands.

The Georgia Tech Research Findings on Local AI Assistants

Local AI assistants process queries directly on the host hardware rather than routing user interactions through remote cloud server infrastructure. While these local models operate, fragments of user dialogues persist inside the device’s volatile memory allocations even after a user executes a deletion request. To examine this data retention behavior, researchers at Georgia Tech constructed a software utility called ORISA designed to scan device memory structures for residual chat fragments.

During experimental trials, the team built a simulated AI application named “Project Note Keeper” utilizing the ChatterUI Android application environment. Investigators loaded the test assistant with 100 distinct factual statements concerning a mock project before issuing a deletion command to clear the session history.

ORISA Tool Retrieves More Facts than Active Memory

Following the deletion procedure, the ORISA tool successfully extracted information from the purged conversation and reconstructed a functional working session. When investigators prompted the restored assistant to recall the project details, the software retrieved 26 distinct facts.

By comparison, the original active session had access to only 14 facts during standard operation because the remaining data points had previously fallen outside its immediate active memory threshold. The experiment demonstrated that recovery utilities can harvest information that the AI model had stopped actively processing.

“It can help establish what information someone shared with an AI assistant and what assistance they received,” stated Ken Xu, a PhD student and lead author of the study.

Scope of Testing Across Android Applications and AI Models

The research team evaluated the ORISA recovery framework across 10 different Android applications paired with three distinct AI models. Across the 60 total test configurations, the tool successfully retrieved conversation data that remained resident in device memory until system activity overwrote the storage blocks.

The findings indicate that residual chat information diminishes and eventually becomes unrecoverable only as the device allocates memory to other running processes and system operations overwrite the legacy sectors.

More on Android: Google unveils new Android security tools to block phone scams · Samsung Android XR Smart Glasses Leak: Specs, Styles, and Shipping Date