Research Tool Recovers Deleted On‑Device AI Chats from Android Phones, Highlighting Logging Gaps
What Happened – Researchers at Georgia Tech released ORISA, a forensic utility that scans an Android device’s working memory to reconstruct portions of conversations that were deleted from a locally‑run AI assistant. In their experiment, the tool recovered 26 of 100 fictional facts that had been removed from the chat history of a test app (ChatterUI). The technique leverages residual batch‑token and key‑value caches that remain in process memory after a session ends. The authors argue that the lack of centralized logging for on‑device AI creates a blind spot for both defenders and investigators.
Why It Matters for Trust & Control Assurance – The ability to retrieve deleted AI interactions underscores a control gap: without systematic logging and audit trails, organizations cannot prove what AI‑driven advice was given or verify that malicious use was not concealed.
Who Is Affected – Mobile‑first enterprises, developers of on‑device AI assistants, and any organization that relies on local AI for confidential workflows (e.g., healthcare, finance, legal).
Recommended Actions –
- Deploy endpoint logging that captures AI assistant API calls and session metadata.
- Incorporate memory‑sanitization routines or secure enclaves to purge residual data after a session.
- Update data‑retention policies to cover on‑device AI artifacts and align them with audit requirements.
Technical Notes – ORISA extracts token sequences from the KV cache of the AI process, then reassembles them into a usable context. No CVE is involved; the issue is architectural – local AI lacks a centralized audit log, leaving forensic evidence in volatile memory. Source: ZDNet article