Moonshot AI Releases Kimi K3.1 Pro with 1M Token Context and Open Weights
Moonshot AI released Kimi K3.1 Pro, a 3T-parameter open-weights model with a 1 million token context window, alongside the 2-3x faster Kimi K3.1 Pro Highspeed; the same day MiniMax open-weighted M3.1 Flash with 428B total parameters and 23B active per token. OpenAI detailed the hardening and red-teaming of its frontier model training environment, while Anthropic published a Claude Mythos 5 alignment assessment and the credential isolation design of the Claude Duet agent.
Model Releases & Updates
Moonshot AI Releases Kimi K3.1 Pro with 1M Token Context; Open Weights to Be Released Before November 4
Moonshot AI released Kimi K3.1 Pro, the first model in the Kimi K3.1 series, a 3T-class open-weights model with a 1 million token context window, and simultaneously launched Kimi K3.1 Pro Highspeed, running 2-3 times faster.
Source：Moonshot AI: Kimi Blog
MiniMax Open-Weights M3.1 Flash: 428B Total Parameters, 23B Active per Token
MiniMax announced that M3.1 Flash weights are public, with 428B total parameters and 23B active parameters per token, available for download on Hugging Face and ModelScope.
Source：X: MiniMax (@MiniMax_AI)
Product Releases & Updates
Kling 4.0 Launches on fal with Open API Access for All Customers
fal announced that Kling 4.0 is now generally available to the public, with API access for every customer. The model supports native video generation of up to 30 seconds with audio, up to 15 image/video/element references, and precise control over up to 10 keyframes, with improved motion, character consistency, and cinematic narrative performance.
Source：X: fal (@fal)
Hesai Launches the Kosmo 3D World Data Engine and the 8K AI Spatial Camera Kosmo One
On October 14, Hesai officially released the Kosmo 3D world data engine, together with the AI spatial camera Kosmo One, which integrates an 8K imaging system, and the cloud platform Kosmo World, targeting real-world data demand for embodied intelligence and world models.
Source：Hesai
Industry News
OpenAI Details the Hardening and Red-Teaming of Its Frontier Model Training Environment
OpenAI described the security hardening of its frontier model training and evaluation environments, including no internet access during training, building an offline cached web world, layered isolation with defense in depth, and assembling teams of the strongest models (including the upcoming GPT-6 Astra Cyber) for high-compute red-team testing.
Source：OpenAI: official news feed (RSS · excludes enterprise/customer cases)
iPhone Duo Goes on Sale at Apple Stores Worldwide, With the First Foldable iPhone on the Shelf
On October 23, iPhone Duo went on sale at Apple Store locations worldwide, where customers can try the first foldable iPhone in person.
Source：Apple: Newsroom (RSS)
State Council General Office Issues Guidance to Amplify AI's Job-Creation Benefits and Ease Job Displacement
On October 23, the General Office of the State Council issued Guiding Opinions on Adapting to the Development of AI Technology to Help Build an Employment-Friendly Model of Development, calling for strengthening AI's positive effects on job creation and the empowerment of traditional roles while effectively easing the job-displacement impact of AI adoption.
Source：ITHome: Artificial Intelligence
Latent Space Maps Out the PHP2Shell Incident: OpenAI's Patch the Planet and Wiz Each Found PHP Core Vulnerabilities
Latent Space's AINews compiled the disclosure timeline of the PHP core vulnerabilities PHP2Shell (CVE-2026-109618 and CVE-2026-109619). Researchers found the flaws in the PHP engine's memory overflow and fatal error handling, and demonstrated pre-authentication RCE and remote memory leaks against WordPress, Drupal, and Joomla.
Source：Latent Space (RSS)
WSJ: OpenAI Fires 3 Security Researchers Over Burner Signal Chats
According to WSJ, OpenAI security personnel fired three security researchers and one external security researcher after discovering they had been using Signal chats with disappearing messages.
Source：X: Rohan Paul (@rohanpaul_ai)
Research & Papers
Anthropic Analyzes Misaligned Behavior of Claude Mythos 5 in the UK AISI CTF Test
Anthropic published an assessment report on the alignment problems of Claude Mythos 5, focusing on misaligned behavior in offensive and defensive benchmark tests.
Source：Anthropic: Research (publications · web)
Anthropic Explains the Safety, Privacy, and Credential Isolation Design of the Claude Duet Agent
Anthropic described the safety and privacy design of the Claude Duet agent, which is in research preview. Important actions such as sending messages, making payments, and deleting data require user confirmation on an approval card, and classifiers independent of the model check whether actions match intent; passwords, cookies, and login tokens never pass through the model, payments use one-time virtual cards, and sensitive information is split into never-share and share-with-consent tiers, checked by an independent classifier before sending.
Source：Claude: Blog (web)
Tips & Opinions
Sundar Pichai Explains Demis Hassabis Stepping Back from DeepMind Day-to-Day Management and Jeff Dean's Departure
In an interview, Sundar Pichai explained why Demis Hassabis is gradually stepping back from running DeepMind day to day and why Jeff Dean is leaving Google, saying that frontier model development has become a large-scale undertaking that consumes everything from data and compute to team management.
Source：X: Alex Heath (@alexeheath)
Quick Brief
- OpenAI 下调估值目标至 1.25 万亿美元，拟融资至多 300 亿美元（The Decoder：AI News（RSS））
- Uber 向全体员工部署 ChatGPT Work（OpenAI：官网动态（RSS · 排除企业/客户案例））
- Grok Bot 新增保存登录与支付信息（X：Grok Bot (@bot)）
【Hacker News Top Stories】
Keywords：AI OR GPT OR LLM OR Claude OR OpenAI OR "machine learning" OR "large language model" OR AGI OR "generative AI" OR AI agent OR RAG OR diffusion OR transformer
Source: hnrss.org | Curated from Hacker News
1. Evolving Threat: http2shell is an RCE in node, Python, and more
Article URL: https://nitter.click/AikidoSecurity/status/2113759110625906950 Points: 5, Comments: 2
2. Securing our frontier model training environment
Article URL: https://openai.com/index/securing-our-frontier-model-training-environment/ Points: 63, Comments: 74
3. Show HN: Corgi, a helpful {agent, terminal} multiplexer
Article URL: https://corgi.computer/ Points: 14, Comments: 1
4. Dessa: the most authentic full-duplex speech model ever built
Article URL: https://block.xyz/dessa Points: 186, Comments: 80
5. Kimi K3.1 Pro and K3.1 Pro Highspeed
Article URL: https://www.kimi.ai/blog/kimi-k3-1-pro Points: 658, Comments: 420
First published on the WeChat official account "瑞哥观势".
About the Author
ERIC
AI Technology Expert, focusing on research and application of artificial intelligence and automation tools
Contact & Platforms