AI Daily: Microsoft Copilot Cowork, OpenAI Security Strategy, and Tencent SongGeneration 2 Music Model

Beyond Just Chatting! Microsoft Let’s AI Finish Your Work Directly, Plus Claude Code Support and Tencent’s Top-Tier AI Music

Recently, when you open your computer, you can probably feel that artificial intelligence has moved to a whole new level. Honestly, people used to expect AI only to help write drafts or answer a few simple questions. But look at the situation now; tech giants are no longer satisfied with this “Q&A game.” The current trend is clear: AI is gradually transforming from a “talking encyclopedia” into a “digital employee who can truly roll up their sleeves and get things done.”

This article will give you an overview of the latest major tech updates. From Microsoft’s new office mode to OpenAI’s security strategy, from the developer-favorite code review tools to music models capable of generating broadcast-quality songs. Ready? Let’s see how these technologies will change our daily work and life.

Let AI Work for You Directly: Copilot Cowork Debuts

Guess what? The long-awaited “office automation” has finally taken a huge leap forward. Microsoft recently officially announced Copilot Cowork: A new way of getting work done. The original design intent of this new feature is very direct: to let Copilot take actual actions rather than just chatting with you.

Imagine this scenario. Monday morning, your inbox is full of customer replies, and your calendar is a mess. At this point, just give Cowork a clear goal, and it will automatically understand the context based on emails, meeting minutes, and files. This relies on Work IQ technology, which can collect signals across software like Outlook, Teams, and Excel and transform user requests into specific execution plans.

Many people might ask, is it really safe to leave it to AI? Will it send emails randomly? Honestly, Microsoft has already thought of this. Cowork executes the plan silently in the background and sets clear checkpoints. It will recommend the next action to you, and everything must be approved by you before it is actually implemented. This mode allows users to advance a dozen tasks simultaneously while never losing control. It’s really smart and practical.

Strict Oversight: OpenAI Acquires Promptfoo to Strengthen System Security

As AI becomes smarter and begins to access core corporate data, security naturally becomes the biggest consideration. OpenAI is clearly aware of this pain point. They just announced that OpenAI is preparing to acquire Promptfoo, a security platform specifically designed to help companies identify and fix AI vulnerabilities during the development phase.

This acquisition is a huge reassurance for enterprise users. As more AI assistants are deployed into real workflows, companies urgently need systematic methods to test agent behavior. The Promptfoo team, led by Ian Webster and Michael D’Angelo, previously built a powerful open-source command-line interface and library specifically for “red-teaming” large language models—simulating hacker attacks to find system weaknesses.

Once the acquisition is complete, this technology will be directly integrated into the OpenAI Frontier platform. Some wonder if the original open-source project will be shut down. Actually, it’s quite the opposite; OpenAI has promised to continue maintaining this open-source project, which is definitely good news for the entire developer community.

Super Inspector for Code: Claude Code Review

After talking about office and security, let’s look at something for developers. The new feature launched by Anthropic is almost like equipping every development team with a tireless senior engineer. You can check out this post Bringing Code Review to Claude Code, which details this agent-team-based code review system.

The way this system works is very interesting. Whenever someone opens a Pull Request (PR), the system sends out an AI agent team. These agents search for errors in the code simultaneously, verify each other to filter out false positives, and finally sort them by severity. This is a mechanism built for thorough inspection, pursuing accuracy and not taking a superficial scanning approach.

According to actual tests, each review takes about 20 minutes on average. While this is slightly more expensive than lightweight solutions—averaging about $15 to $25 per review—it can catch hidden errors that human reviewers often miss. You might wonder if this system will automatically approve PRs? Don’t worry, the final decision still rests with human engineers. This tool is only responsible for closing loopholes so developers can focus more on the core logic of the code.

Finding the Coolest Tools: OpenRouter’s New Rankings

There are so many developer tools that sometimes it’s really hard to know which one to use. To make the ecosystem more transparent, OpenRouter has launched its new App & Agent Rankings. This is a beacon for users who want to discover potential applications.

This ranking doesn’t just look at total volume; it also allows applications to be ranked in different categories and even lists weekly hot trends. In addition to traditional token usage, ranking criteria now include the number of requests. Developers can also easily opt-in and categorize their generation results. This way, everyone can see very intuitively which AI applications are currently the most popular in the market.

A Surprise for Reading Enthusiasts: NotebookLM Supports ePub

Changing the mood, let’s look at an update closely related to daily learning. If you like to bury your head in books at the library, this news will surely excite you. NotebookLM officially released a great announcement on social media, announcing that they finally support uploading ePub format files.

This is definitely one of the most requested features by users. Now, you can easily upload e-books, study guides, or even long novels into the system, making these static words instantly “come alive.” You can directly ask the AI about plots in the book, request summaries of key knowledge points, or even discuss specific chapters. Learning efficiency is thus greatly improved.

Open Source Music Approaching Commercial Standards: Tencent SongGeneration 2

Finally, let’s talk about an amazing multimedia technology breakthrough. SongGeneration 2, jointly developed by Tencent and the Tsinghua University Human-Computer Speech Interaction Lab, has officially arrived. This model has dropped a bombshell in the field of music generation.

Its biggest highlight is the “dual-core” hybrid architecture with division of labor. The language model acts as the “composing brain,” responsible for arranging the overall musical structure. The diffusion model transforms into a “high-fidelity renderer,” responsible for synthesizing extremely complex acoustic details. Additionally, it pioneered a layered representation structure, including hybrid and multi-track representations, allowing precise control over subtle changes in vocals and accompaniment.

Various evaluation data show that SongGeneration 2 exceeds many mainstream models in several dimensions such as overall quality, melody, and arrangement, with extremely low phoneme error rates in lyrics. If you want to experience it yourself, go to their Demo Page to hear the effects. Developers can also get the “large” model with 4B parameters directly from the Github page or HuggingFace page. For those pursuing generation speed, there is also a Fast version on the HuggingFace Space platform that can generate a complete song in under a minute.

However, there is a very critical note here. Some might ask, can such a powerful model be used to make YouTube videos and turn on monetization? The answer is absolutely not. The licensing terms are very strict, clearly stating that it is for academic, research, and educational purposes only, and any form of commercial or production environment use is strictly prohibited. This restriction covers both the inference code and the model weights. Users must pay special attention to this before using it.

The evolution of technology is always dazzling. Whether it’s improving office efficiency, overseeing system security, or igniting creative sparks, these new tools are truly changing the way we solve problems. Go try these new features; maybe you’ll find unexpected surprises.