https://guijiagi.com/categories/%E5%AE%89%E5%85%A8/Recent content in 安全 on AI 实战派 · 从技术到赚钱Hugozh-cn本站内容采用 CC BY-NC-SA 4.0 国际许可协议授权Wed, 07 Oct 2026 17:42:00 +0800https://guijiagi.com/posts/2026-10-07-ai-model-supply-chain-attack/Wed, 07 Oct 2026 17:42:00 +0800https://guijiagi.com/posts/2026-10-07-ai-model-supply-chain-attack/PyTorch pickle 反序列化漏洞、投毒模型、被篡改的 LoRA 适配器。拆解 AI 时代的供应链风险与防御底线。https://guijiagi.com/posts/2026-10-07-training-data-extraction-leakage/Wed, 07 Oct 2026 17:36:00 +0800https://guijiagi.com/posts/2026-10-07-training-data-extraction-leakage/Carlini 等人证明大模型会逐字复述训练语料。拆解记忆与泛化的边界，以及隐私数据如何从权重里被「钓」出来。https://guijiagi.com/posts/2026-10-07-llm-jailbreak-gcg-autodan/Wed, 07 Oct 2026 17:30:00 +0800https://guijiagi.com/posts/2026-10-07-llm-jailbreak-gcg-autodan/GCG、AutoDAN 这类自动越狱方法用优化找出模型看不见的后缀。拆解越狱的攻击机理与对齐研究的军备竞赛。https://guijiagi.com/posts/2026-10-07-prompt-injection-attack-defense/Wed, 07 Oct 2026 17:24:00 +0800https://guijiagi.com/posts/2026-10-07-prompt-injection-attack-defense/间接提示注入能让检索到的网页、文件内容劫持 Agent。拆解攻击路径、为什么难防，以及纵深防御的现实手段。