AI 日报 2026-11-07

【今日速览】

今天AI界有点热闹，不仅有新模型发布，还有开源项目不断涌现。OpenAI发布了GPT-6 Cyber，性能提升明显；同时，开源社区也热闹非凡，各种有趣的项目让人眼前一亮。比如，一个专门为书籍推荐的网站HackerBooks，还有让AI在iPhone上操作的poki项目，都挺有意思的。

Sebastian Raschka的博客上发布了一系列关于LLM架构和推理优化的文章，包括KV Sharing、mHC和Compressed Attention等（HN 821分·312评论）。

Parameter Labs描述了他们发现的ChatGPT实现中的artifacts模式漏洞，并展示了如何利用该漏洞接管ChatGPT账户（HN 821分·312评论）。

今天AI界热点集中在模型发布和开源项目上，OpenAI的GPT-6 Cyber和Ant Ling的Ling-3.2-flash都值得关注。同时，开源社区也活跃异常，各种有趣的项目不断涌现。总的来说，今天AI界充满了活力和创造力。

Show HN: HackerBooks – HackerTube, but for books mentioned on HN (1864pts·0评)
Concerning behavior from a popular open-weight AI model (337pts·0评)
Launch HN: Rivet (YC W23) – Scale and Orchestrate Agents on S3 (118pts·0评)
Malware uses LLMs to rank victims by how much they're worth (263pts·0评)
ChatGPT and Codex usage limits merging (236pts·0评)
Launch HN: Kino AI (YC S23) – AI media manager and video editor (109pts·0评)
Launch HN: Momentic (YC W24) – Agents that catch bugs by using your app (166pts·0评)

[DialFork: a benchmark for procedural hallucination in conversational agents [R]](https://www.reddit.com/r/MachineLearning/comments/1xofjrn/dialfork_a_benchmark_for_procedural_hallucination/)
No GPT-6.2, no Fable 5.5, no Qwen 4
Finally coming. Qwhen4 tease friday is here
OpenAI has just announced GPT‑6 Cyber, a major step up from GPT‑5.6 Cyber and outperforms GPT‑6 Astra on most cyber tasks
If the custom GPT you use every day was made by someone else, you can't migrate it before Dec 11. OpenAI's FAQ says your access doesn't carry over to the replacement either
[Ant Ling introduces Ling-3.2-flash: 560B/25B native multimodal MoE with 1M context [N]](https://www.reddit.com/r/MachineLearning/comments/1xpye7g/ant_ling_introduces_ling32flash_560b25b_native/)