Mirage bounds LLM KV-cache timing leaks without blocking sharing
On a 500-sample corpus and three models, token-emission delays pushed leakage near random guessing with lower TTFT than baselines.
Stories per day · 30 days
1 stories, newest first · one card per story