爱窝啦 AI 日报

Today’s Summary

OpenAI urgently halted GPT-6.1 after testing revealed zero-day vulnerabilities in mainstream browsers, while Anthropic's tests showed Zhipu's GLM-5.3 can independently write attack programs.
Product launches, open-source tools, and research papers are all tackling the same issue: how to prevent AI from losing control after gaining system privileges, from sandbox environments to scoring system calibrations—everyone's catching up.
Today, focus on the top three security incidents first. Teams that have deployed open-source models or granted Agent permissions need to redo their risk assessments immediately.

🔥 Today’s Top 10 Highlights

1. OpenAI Halts GPT-6.1, Testing Revealed It Could Crack Mainstream Browsers

Launch plan completely withdrawn. GPT-6.1 Astra, originally scheduled for October release, has been urgently halted , according to 36Kr. Testing showed the model discovered multiple zero-day vulnerabilities 🔓 in mainstream browser JavaScript engines within a single day, chaining them into attack webpages capable of reading users’ arbitrary files. These vulnerabilities have been reported to maintainers. The withdrawal on the eve of the developer conference signals that OpenAI is tightening its security review standards.

2. Anthropic Tests Show Zhipu’s GLM-5.3 Can Independently Write Attack Programs

Open-source model security defenses bypassed. GLM-5.3, Anthropic’s tests show Zhipu’s GLM-5.3 can independently write attack programs , according to Baoyu’s report, performing at a level close to Claude Mythos Preview, which is only available to select institutions. Researchers used it to discover multiple zero-day vulnerabilities 🔒 in a browser within one day, with only 20 minutes of human input. Simple methods could bypass its security protections. Teams currently deploying open-source models need to conduct an additional security assessment.

3. NVIDIA Open-Sources OpenShell to Provide Sandbox Environment for Autonomous AI

Agents now have independent virtual machines. OpenShell, NVIDIA open-sources OpenShell to provide sandbox environment for autonomous AI , has reached 10,606 total stars. This Rust-written secure runtime allows AI agents to execute commands and install tools in isolated environments 💻 without affecting the host system. Suitable for teams developing Agent products that require system-level operational privileges.

4. Open-Source Bastion Host JumpServer Enters V5 Era

Privileged access management platform upgraded. JumpServer, open-source bastion host JumpServer enters V5 era , according to a GitHub roundup. This enterprise-oriented open-source bastion host can integrate Linux servers, databases, Kubernetes clusters, and internal management backends, centrally managing access permissions. Small and medium-sized teams needing centralized auditing of developer and operations activities can use it as an alternative to commercial solutions.

5. Instinct Personal Assistant Free for Users, Charges Merchants Transaction Fees

New commercial model for Agents. Instinct personal assistant, free for users, charges merchants transaction fees , according to Mo Weishu’s roundup. It’s free for users and charges merchants transaction fees 💰. Within 3 weeks, 40% of users share credit cards, with retention rates around 80% after sharing. Typical use cases include wardrobe scanning and styling, online shopping, travel booking, and Uber ridesharing. Teams designing subscription-based AI products can evaluate this approach.

6. Grok Bot/Cue/dots—Three Products Spark Identity Confusion Debate

Same-name products confuse users. The products shown in the first three screenshots, many mistook for being the same product , according to Huajuan’s post on Jike. These three belong to different vendors with varying functions and positioning 🤔. Comments indicate naming conflicts are affecting user identification. When naming AI products, search existing brands first.

7. Agent-Generated Multiplayer Games May Gain Mainstream Adoption Faster Than Office Scenarios

From production to consumption. Comparing Agents to early personal computers is more accurate, benn writes on Jike . Personal computers proved their production value through office software, but truly entered households through multimedia, games, and chat rooms 🎮. Agent-generated multiplayer games don’t rely as heavily on game content—the fun comes more from player interactions, where everyone modifies and plays together. Developers watching Agents transition from office to entertainment can focus on this direction.

8. META Proposes RL-XAR Solution to Address AI Writing’s Empty Content Problem

Teaching judges to recognize good writing first. META proposes RL-XAR solution, to address AI writing’s empty content problem , according to Gorden Sun’s report. The RL-XAR approach first adjusts judges’ standards until they can accurately rate human expert texts as high quality ✍️, then uses these calibrated rules to retrain the writing model. In academic papers, novels, and encyclopedia entry continuations, model output quality improved markedly. Teams training writing models can reference this calibration approach.

9. Guizang Integrates Jimeng CLI into Muse for Image and Video Generation

Agent nesting plan advances further. Guizang demonstrates, he installed Jimeng’s CLI tool in Muse’s virtual machine , and now it can help him draw images and generate videos using Seedance 2.5 🎨. Previously, he had Muse install DeepSeek Harness specifically for querying domestic information. Agents with virtual machines make tool combinations more flexible.

10. Using AI to Generate Surveillance Videos Then Phone Recording Them—Millions of Views on Douyin

Edge tactics harvest traffic. According to AI Exploration Guide channel, someone uses AI to generate surveillance videos of foot spas, KTVs, and billiard halls , then films the result videos with a phone and posts them on Douyin, avoiding AI-generated labels 📹. View counts reached several million. This approach bypasses platform labeling rules but risks immediate bans. Content creators need to weigh traffic against compliance risks.

⌘ Top Open-Source Projects

VectifyAI/PageIndex: Reasoning-Based Document Indexing Without Vectors

New option for RAG systems. PageIndex gained 835 Stars in one day , reaching 37,373 total stars. It uses reasoning instead of traditional vector retrieval, indexing documents without relying on embedding models 📑. Suitable for teams optimizing RAG recall effectiveness and hoping to reduce vector database dependencies.

😄 AI Fun Facts

Boss Drops a Video Release Request at 9 PM

Manus Studio reads meeting minutes, Notion, Slack, Figma and various other materials , stringing them together as an inspiration mentor to help the blogger refine storyboards. The video finished in half a day not only matched brand tone but also scored the blogger’s second-ever million views on X—now they’re already using it to edit vlogs. AI doesn’t replace creativity, it just amplifies one person’s creativity a hundredfold.