ORISA tool extracts deleted AI chats from Android memory

What ORISA can recover

Researchers at Georgia Tech, led by Ph.D. student Ken Xu, developed ORISA, an experimental tool designed to recover deleted AI chat sessions from an Android device’s memory. The tool targets local AI applications, which process user queries on-device rather than in the cloud.

In a controlled test, ORISA reconstructed 26 fictional facts from an original set of 100, surpassing the 14 facts the original AI assistant session had access to. This suggests that deleted or transient data may persist in memory longer than expected.

How the tool works

ORISA operates by analyzing process memory dumps from Android devices. It reconstructs token sequences and session-level attention states left behind by AI applications, assembling fragments into usable data. The researchers tested the tool using ChatterUI, a framework for local AI chatbots.

The recovery process is time-sensitive. ORISA’s efficacy diminishes as the device’s memory is overwritten by newer activity, but it can still retrieve partial or complete sessions shortly after deletion.

Why local AI security gaps matter

Unlike cloud-based AI, local AI lacks centralized logging, enforcement, or audit trails. This makes it difficult to verify whether sensitive data has been securely deleted or if malicious actors have tampered with the model.

The research highlights a broader risk: bad actors could use tools like ORISA to recover deleted queries and fine-tune models for malicious purposes, such as phishing or fraud, without leaving traces.

For users and operators, this means sharing sensitive information with local AI chatbots carries inherent risks, even if the data is deleted after use.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

ORISA’s findings expose a tangible risk for founders and operators building or deploying local AI applications on consumer devices. While cloud-based AI services can enforce strict data retention and deletion policies, local AI runs entirely on-device, leaving memory artifacts that tools like ORISA can reconstruct. This means sensitive user queries—even those deleted—may still be recoverable, creating potential compliance and reputational risks.

For teams focused on privacy-preserving AI, this research reinforces the need for memory sanitization techniques or hardware-level protections to prevent data leakage. It also signals to regulators and enterprise buyers that local AI is not inherently more secure than cloud alternatives—merely harder to audit.

Key takeaways

• Local AI applications on Android may leave recoverable traces of deleted user interactions, posing privacy and compliance risks.
• Tools like ORISA can exploit unmanaged process memory, making it harder to guarantee data deletion.
• Founders relying on local AI for sensitive use cases must address memory forensics risks or risk regulatory scrutiny.

Related on Lazyfounder

• More technology news