What happened
Georgia Tech researchers built ORISA, a forensic utility that scans an Android phone’s working memory for remnants of deleted AI chat sessions. In a controlled test using the ChatterUI app, ORISA recovered 26 of 100 fictional facts that had been removed from the chat history. The technique relies on residual batch‑token and key‑value caches that persist in process memory after a session ends. The authors note that on‑device AI lacks centralized logging, making it difficult for investigators to trace malicious use or verify compliance.
Why it matters for trust and compliance
- The finding highlights the need for systematic logging of AI assistant activity on endpoints to satisfy audit and forensic requirements across frameworks.
- Demonstrates a control gap in AI logging that maps to audit‑ready evidence requirements.
- Supports continuous monitoring of AI‑driven processes to meet multiple framework objectives.
Who is affected
Mobile‑first enterprises AI assistant developers Organizations using on‑device AI for confidential tasks
Recommended actions
- Implement endpoint logging that captures AI assistant API calls and session metadata.
- Add memory‑sanitization or secure enclave mechanisms to purge residual AI data after use.