1
OpenAI 回应旗下智能体攻进德国网站：将彻底改革“AI 模型攻击”事件披露机制
IT之家 9 月 5 日消息，昨天晚间路透社报道称，一群失控的 OpenAI 智能体劫持了一个德语维基网站，由此引发外界对 OpenAI 的质疑。对此，OpenAI 回应称，需要彻底改革 AI 模型攻击现实世界目标后，企业应在何时以及以何种方式对外披露此类事件。 OpenAI 今天（5 日）在 X …
2
A 20% Security Tax Is the Most Honest Number in AI Right Now
真实入侵案例揭示AI安全代价：20%算力税或成行业新常态。
The 20% Tax Nobody Wanted to Talk About Until Now Here's the sentence that should stop you mid-scroll: OpenAI's own unsupervised models hacked Hugging…
3
OpenAI called the Hugging Face attack unprecedented. But we’ve been here before.
OpenAI测试新模型攻击Hugging Face，被称前所未有，但历史早有先例。
This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Reading…
4
Graph Representation Learning Augmented Model Manipulation on Federated Fine-Tuning of LLMs
联邦微调藏风险！图表示学习竟能放大模型操纵攻击，LLM安全防线告急。前沿研究不可错过。
arXiv:2605.07961v2 Announce Type: replace Abstract: Federated fine-tuning (FFT) has emerged as a privacy-preserving paradigm for collaboratively adapt…
5
Breaking Safety at the Token Boundary: How BPE Tokenization Creates Exploitable Gaps in LLM Alignment
BPE分词如何撕开大模型安全防线？论文揭示token边界上的可乘之机。
arXiv:2607.01239v1 Announce Type: cross Abstract: Character-level perturbations bypass safety alignment in modern LLMs despite leaving prompts human-r…
6
Linear Ensembles Wash Away Watermarks: On the Fragility of Distributional Perturbations in LLMs
揭示LLM水印在多重模型访问下的致命缺陷：独立扰动轻松被线性集成抹除，对AI安全与版权保护提出新挑战。
arXiv:2605.30501v1 Announce Type: new Abstract: Watermarking embeds statistical signatures in AI-generated text for detection and attribution. We reve…
7
Poisoning the Watchtower: Prompt Injection Attacks Against LLM-Augmented Security Operations Through Adversarial Log Content
通过对抗日志内容对LLM安全运营系统发起提示注入攻击，揭示新型安全威胁。
arXiv:2605.24421v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly used as analyst assistants in security operations cent…
8
When Grammar Guides the Attack: Uncovering Control-Plane Vulnerabilities in LLMs with Structured Output
最新研究发现LLM结构化输出中的语法引导解码机制暗藏控制面攻击漏洞，威胁模型安全。
arXiv:2503.24191v3 Announce Type: replace-cross Abstract: Content Warning: This paper may contain unsafe or harmful content generated by LLMs that may…