Recurrent states in AI models leak private information through power use
Privacy Leakage from Recurrent State in LLM Inference
Cryptography and Security
Summary
Some AI models remember your private information inside their hidden memory during conversations, even if that information is not shown again. The authors found ways to guess sensitive facts by looking at these hidden states and even by watching the power the computer uses while running the AI. This means that private data can accidentally leak out through how the model keeps and uses its memory during interactions. The paper points out new privacy risks and suggests system builders and hardware teams should be careful about who accesses this hidden data and how the hardware behaves.
What this means in practice
- •For llm service providers: Prevent unauthorized access to recurrent state data to protect user privacy during personalized LLM interactions.
- •For gpu operators: Restrict telemetry and power monitoring to avoid leakage of user information through power usage patterns during AI computations.
Authors
Weimin Fu, Shikun Liu, Qiying Li, Yuntao Liu, Pan Li, Xiaolong Guo
Abstract
Personalized LLM services place private facts and application instructions into a prompt, then retain representations of that context for later inference. Latent-state LLMs compress part of this history into fixed-size recurrent state. The information preserved by its learned updates is implicit, and later computation can expose it even when an observer receives neither the private text nor the responses. We investigate privacy leakage from recurrent state through partial-state probing and physical observation of its use. Our methodology separates recoverable information, persistence through public dialogue, causal influence on later decisions, and leakage through power. Experiments with released hybrid implementations show that a small state observation supports multiple predictions about private facts and response rules, and that personal attributes generalize to unseen authors. The information remains inferable after public dialogue and changes later decision margins. Complete-read power recovers private facts under a fixed execution schedule. On a partitioned GPU, an unprivileged account infers age from whole-device power during public requests; a separate state-isolation experiment establishes the recurrent contribution to this channel. These findings extend context confidentiality to the lifetime and execution of retained state. They identify access-control responsibilities for serving systems, telemetry restrictions for GPU operators, and data-dependent power leakage for accelerator designers.