Side-channel attack reconstructs local LLM outputs from CPU cache

A new paper from Microsoft and others shows that observing CPU cache activity during detokenization can reconstruct locally hosted LLM outputs with up to 95% success, challenging assumptions that local deployments are safe.

2026-09-09 ~ 2026-09-11 · 4 related posts

• New side-channel attack reconstructs local LLM outputs from CPU cache traces — chaumian · 2026-09-09
• Microsoft researchers show CPU cache attack that reconstructs local LLM output via the detokenizer — dair_ai · 2026-09-10
• Local LLMs leak outputs via CPU cache side channel, up to 95% reconstruction success — rohanpaul_ai · 2026-09-11

1 near-duplicate retellings: rohanpaul_ai