你有没有想过，在AI安全的攻防战中，为什么防御者总是慢半拍？我们能否跳过对话，直接把指令“注入”AI的大脑？在众多复杂的AI模型背后，是否存在一个统一所有武功的“心法总纲”？今天的节目，我们将通过几篇最新论文，一同寻找这些问题的答案，甚至尝试给AI的思考过程做一次“脑部CT”，看看它到底是如何想问题的。

00:00:32 AI安全的“纸上谈兵”：为什么说攻击者总是后出手的那个？

00:05:36 AI的“意念注入”：如何把指令直接写进模型大脑？

00:11:22 AI大模型的心法：一个统一所有武功的“总纲”

00:18:58 给大模型装上导航，能不能开得更快？

00:23:38 给AI做个脑CT：看清它思考的脉络

本期介绍的几篇论文：

[LG] The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against LLM Jailbreaks and Prompt Injections

[OpenAI & Anthropic & Google DeepMind]

https://arxiv.org/abs/2510.09023

---

[LG] Transmuting prompts into weights

[Google Research]

https://arxiv.org/abs/2510.08734

---

[LG] Design Principles for Sequence Models via Coefficient Dynamics

[ETH Zurich & ELLIS Institute Tübingen]

https://arxiv.org/abs/2510.09389

---

[LG] The Potential of Second-Order Optimization for LLMs: A Study with Full Gauss-Newton

[Harvard University]

https://arxiv.org/abs/2510.09378

---

[CL] Verifying Chain-of-Thought Reasoning via Its Computational Graph

[FAIR at Meta]

https://arxiv.org/abs/2510.09312