Pai 的 AI 雷达

安全治理

今天：Paul Christiano 加入，Bait-and-Recover（arXiv:2609）。并行还有 Sep 9。

日期: 2026/09/10
今日条目: 78
当前类目: 3

往期2026/09/10

全部78 编码代理0 Agent与工具3 个人助理1 模型与评测26 创作成片11 开发与开源16 产品接入21

Paul Christiano 加入 OpenAI Foundation 董事会

Paul Christiano joins the OpenAI Foundation Board and its Safety and Security Committee, bringing experience in AI alignment, safety, and standards.

openai.com打开原文

#02

Bait-and-Recover: Poisoning Internal Refusal Signals to Defend LLMs against White-Box Editing Jailbreaks

arXiv:2609.05794v1 Announce Type: cross Abstract: Open-weight large language models face a low-cost white-box threat from representation engineering attacks. Attackers can estimate refusal directions and search for projection-matrix edits that suppress safety alignment while pres

arxiv.org打开原文

#03

Sep 9, 2026AlignmentAn alignment assessment of recent cybersecurity incidents

anthropic.com打开原文