AI Daily: MiniMax M2.7, MOSS Lightweight Voice, and OpenAI Certificate Update

OpenAI Releases macOS Security Update and Practical Upgrades for Major AI Tools

Here is a roundup of today’s noteworthy AI information. Content includes OpenAI’s app update for the Axios vulnerability, GitHub Copilot resource allocation adjustments, and the latest technical progress of Claude, MiniMax-M2.7, and MOSS-TTS-Nano. Both developers and daily users will find relevant and practical information here.

As we reach mid-April 2026, the spirit of spring brings a sense of vitality. To be honest, there is quite a bit of information today. From the evolution of underlying models to upgrades in daily document processing and critical security alerts, major tech companies have released significant updates. Let’s organize this important information. Just like tidying up a room, classifying information clearly helps us know how to respond next.

Is Your Mac App Secure? OpenAI’s Emergency Response

Security is often a top priority. Did you know that a third-party developer tool named Axios recently encountered a security vulnerability? This involves a wide-ranging software supply chain attack. Out of caution, OpenAI discovered that during their macOS application signing process, an infected version of Axios was downloaded and executed.

While this might sound alarming, OpenAI has intervened quickly. According to OpenAI’s official statement, there is currently no evidence that any user data was accessed, nor is there any sign of intellectual property theft or software tampering.

This is a precautionary measure. OpenAI is revoking old digital certificates. If you are using the macOS versions of ChatGPT Desktop, Codex App, Codex CLI, or Atlas, please be sure to update. After May 8, 2026, older versions of these applications will no longer be supported and may even stop functioning.

Many might ask: Do I need to change my password? The answer is absolutely not. API keys are also unharmed. For readers using Windows, Linux, or iOS, this incident only affects macOS users; web versions and apps on other platforms can still be used safely. Please ensure you download updates through official channels and do not click on installation links from unknown sources.

Developer Daily Adjustments: GitHub Copilot Resource Reallocation

Moving from security to tools developers use every day: as the number of GitHub Copilot users has surged, the system has faced extremely high concurrent requests and usage pressure. This is normal, as everyone wants to use a helpful tool.

To address server resource allocation issues, GitHub released an official announcement declaring new usage limits. The system will now control usage based on overall service stability and specific model families.

One of the most significant changes is for GitHub Copilot Pro+ users: starting today, the Opus 4.6 Fast model service will be discontinued. Users are encouraged to switch to the standard Opus 4.6, which has similar capabilities, or rely on “Auto mode” to distribute request pressure. Think of it like a popular restaurant limiting dining time during peak hours—the goal is to ensure every customer receives stable and fast service. If you have higher requirements, upgrading plans is also an option to consider.

Making Documentation Easier: Claude Officially Arrives in Word

This next update will surely please office workers who frequently write proposals or reports. Anthropic has introduced the Claude for Word feature. That’s right—Claude is now built directly into your Word documents.

In the past, we often had to switch back and forth between browsers and documents. Now, the situation is different. You just need to select a piece of text in Word and tell Claude how you want to modify it; it will present the results using “Track Changes.” This means you can clearly see every addition or deletion and retain final decision-making power.

Even more interesting, Claude can perfectly inherit the document’s original heading styles and numbering logic. It can help you check for inconsistent terminology across the context and maintain the same conversation thread across Word, PowerPoint, and Excel. This seamless experience truly saves a lot of effort in tedious paperwork. It also meets enterprise-grade security standards and complies with existing regulatory frameworks.

Hidden Powerhouse: MiniMax-M2.7 Demonstrates Autonomous Evolution

Shifting our focus from applications to large language models, the MiniMax-M2.7 model page, which features 229 billion parameters, has sparked significant discussion. Its biggest highlight is its strong capability for self-learning and correction.

In internal testing, M2.7 was able to autonomously analyze code errors, modify them, and execute evaluations. After hundreds of rounds of automatic optimization, its performance improved by a full 30%. This “self-driving” evolution process is quite impressive.

In terms of actual engineering performance, M2.7 matches the level of GPT-5.3-Codex. It doesn’t just generate code; it also performs system-level logical reasoning, such as analyzing logs and troubleshooting errors. In multiple simulations, it even reduced server failure recovery time to under three minutes. Furthermore, it has performed excellently on various test leaderboards, surpassing many open-source predecessors. If you’re interested in its real-time visual interaction with users, you can also check out the official OpenRoom interactive demo.

Compact Yet Professional: MOSS-TTS-Nano Speech Generation Model

Finally, let’s talk about a small but beautiful open-source project. For developers who want to add voice functionality to their applications but are limited by hardware resources, this is definitely good news.

The open-source community has introduced MOSS-TTS-Nano. You can hear the actual results at the official demo. If you want to know the technical details, you can visit the GitHub project or download the model directly from the Hugging Face repository.

This speech generation model has only about 100 million parameters, making it perfect for running directly on a CPU without the need for expensive graphics cards. Despite its small size, it supports 20 languages, including Chinese, English, Japanese, and Korean, and can output high-quality 48 kHz stereo audio. Best of all, it can perform voice cloning with just a short reference audio clip. This configuration significantly lowers the hardware threshold while retaining high practicality, making it ideal for building lightweight applications that require real-time voice responses.

In summary, today’s key takeaways are to protect your development tools while not hesitating to try these fresh technologies. The evolution of information technology is always full of surprises; as long as we stay curious, we can find secrets to improving efficiency.

Q&A

🔒 Security: OpenAI macOS Application Vulnerability

Q: What security issue happened with OpenAI’s macOS app? Do I need to change my password? A: During the signing process for its macOS apps, OpenAI downloaded and executed an infected version of the third-party developer tool Axios (version 1.14.1), which was hit by a widespread software supply chain attack. This put the certificates used to sign macOS apps (such as ChatGPT Desktop and Codex App) at risk. As a precaution, OpenAI has revoked that certificate. Affected older macOS apps will stop being supported after May 8, 2026, and may stop working. However, official investigations confirmed that no user data was accessed, and passwords and API keys were not affected, so you do not need to change your password.

💻 Developer Tools: GitHub Copilot Resource Limits

Q: Why can GitHub Copilot Pro+ users no longer use the Opus 4.6 Fast model? A: As usage of GitHub Copilot has grown rapidly, the system has faced extreme concurrent requests and infrastructure pressure. To ensure a fast and reliable experience for all users, GitHub announced new capacity limits and, as a first step in streamlining model offerings, stopped providing the Opus 4.6 Fast model to Pro+ users starting April 10, 2026. Official advice is to use the standard Opus 4.6 or rely on “Auto mode” to switch models.

📝 Daily Productivity: Claude for Word

Q: When Claude for Word is integrated, will it ruin my carefully set layout? A: Not at all. Claude can perfectly inherit original heading styles, numbering logic, and defined terminology. When you select a paragraph and request changes, Claude’s edits appear in Word’s native “Track Changes” mode, allowing you to accept or reject them in the Reviewing Pane. Additionally, it maintains the same context across Word, PowerPoint, and Excel, and can even save common workflows (like contract reviews) as team-shared “Skills.”

🧠 Underlying Models: MiniMax-M2.7’s Self-Evolution

Q: MiniMax-M2.7 claims to have “self-evolution” capabilities. What are the specific results of this? A: MiniMax-M2.7 is a large language model with 229 billion (229B) parameters that can autonomously build complex agent environments to complete tasks. During internal development, M2.7 underwent hundreds of rounds of auto-optimization—analyzing error trajectories, modifying code, and executing evaluations—improving performance by 30%. In system-level reasoning and engineering, it can even reduce real-time recovery time for server failures to under three minutes in simulations.

🗣️ Speech Generation: MOSS-TTS-Nano Lightweight Model

Q: If I have limited hardware (e.g., only a CPU), is MOSS-TTS-Nano suitable for developing multilingual real-time voice features? A: Absolutely. MOSS-TTS-Nano has only about 100 million (0.1B) parameters and is designed to run smoothly on a 4-core CPU without a GPU for streaming voice generation. Despite its size, it supports 20 languages including Chinese, English, Japanese, and Korean, and outputs high-quality 48 kHz stereo audio. Furthermore, it only needs a short reference audio clip for direct Voice Cloning.