本地部署也不安全：CPU 缓存侧信道可还原 LLM 输出，成功率最高 95%
rohanpaul_ai · x · 2026-09-11
TL;DR: A locally hosted LLM can leak its answers through CPU cache activity during detokenization — the routine step of turning tokens into readable text leaves repeatable cache patterns that another local process can learn to reconstruct responses.
- Full-response reconstruction success: 56%–93% on text tasks, 95.87% in one code setting, 30.12% end-to-end against OpenClaw
- Attack requires the attacker to already be on the same machine, sharing CPU resources and profiling the same long-lived LLM process
- Sensitive deployments should keep untrusted software off the same physical core; paper recommends stronger CPU isolation, shorter-lived processes, and disabling SMT where justified
Off-cloud deployment alone is not a complete privacy guarantee.
所属事件:新侧信道攻击可从 CPU 缓存还原本地 LLM 输出(4 条相关)→
「安全」频道最新
- 用 Claude Code 当 iPhone 取证器:挖出「已删」照片的 30 天隐藏副本 — alex_verem · 2026-10-27
- Anthropic 自曝：每 9 个月封禁约占营收 5-7% 的客户 — Polymarket · 2026-10-27
- AVERI 独立审计 Amazon 内部 AI 安全评测，Amazon 据此改进测试流程 — Miles_Brundage · 2026-10-27
- 美中欧 AI 路线分化：美国押注 AGI 霸权，其余世界当公共品 — BubblyExtent1693 · 2026-10-27
- Agent Field Kit 上线：为生产级 Agent 提供 $29 起的安全应急手册 — alvelda · 2026-10-27
- DeepSeek 安全评测发现「过度归因」：模型把普通任务当成考试 — Much_Boysenberry9124 · 2026-10-27