Microsoft researchers show CPU cache attack that reconstructs local LLM output via the detokenizer

dair_ai · x · 2026-09-10

dair-ai highlights a wild new paper from Microsoft and colleagues demonstrating a side-channel attack that reconstructs text a local LLM generates by watching CPU cache activity during detokenization. Unlike earlier cache attacks requiring unusual deployment setups (shared memory, CPU offloading, or MoE), this targets the detokenizer present in default inference pipelines. The two-stage method uses Flush+Reload on shared tokenizer code to detect decoding, fires Prime+Probe to isolate token-dependent cache activity, then recovers readable text via clustering plus a language model. Evaluated across datasets, hardware, frameworks and model families, including real local and agentic deployments.

Related event: Side-Channel Attack Reconstructs Local LLM Output from CPU Cache(2 posts)→

Original post →

More from Safety

Safety channel →