AI Daily | Apple Sues OpenAI for Trade Secret Theft, Grok CLI Privacy Controversy, Claude Fable 5 Highly Rated

AI Daily | Apple Sues OpenAI for Trade Secret Theft, Grok CLI Privacy Controversy, Claude Fable 5 Highly Rated

Apple has formally sued OpenAI for allegedly stealing hardware trade secrets, sparking major concern in the AI industry. xAI has opened Grok 4.5 and Grok Build for free, but the Grok CLI has also triggered privacy concerns over the uploading of entire code repositories. Additionally, this summary covers Claude Fable 5 receiving high praise from Cognition, Google AI Studio launching free custom URLs, the MuScriptor open-source music transcription model, the LingBot-VA 2.0 robot control model, and the latest AI news and technical updates on OpenAI ChatGPT Work, Codex, and more.

Grok 4.5 and Grok Build Open for Free Trial

Honestly, xAI’s latest move is quite striking. Grok 4.5 is now officially open to free users. Anyone with an X or Grok account can directly experience the powerful features of Grok Build. At first glance, this looks like a simple promotional strategy. However, looking deeper, this decision will undoubtedly attract a large number of new users and might even change the market landscape. High-end models that were previously locked behind paywalls are now available to try for free. Developers who want to test the potential of the latest language models should not miss this opportunity.

Grok 4.5 is now available to try on the free tier. Use Grok Build with any X or Grok account.

We’re excited to hear your feedback.https://t.co/NYsa0Ar9eo pic.twitter.com/5H0C2kgrEI

— Grok (@grok) July 10, 2026

Privacy Concerns? Grok CLI Accused of Secretly Uploading Entire Codebases

However, new tools always bring new challenges. Just as everyone was cheering for Grok, a different voice emerged from the cybersecurity field. According to a Cereblab investigation report, when the Grok Build CLI runs, it appears to transmit the user’s entire code repository to xAI’s cloud. Many might ask: what exactly does this include? The answer is all tracked files, even your git history that you thought you had deleted. The most worrying part is that even if the “improve models” option is manually turned off, the uploading action does not stop. This has caused quite a stir in the developer community. Everyone should be extra cautious with data security settings before linking projects to new tools.

MuScriptor: A Major Breakthrough in Music Transcription Technology

Next, let’s look at some relaxing and creative news. The open-source music transcription model MuScriptor, jointly developed by the Kyutai and Mirelo teams, was just launched, and this technology is truly eye-catching. Just input an audio recording, whether it’s pop, classical, or heavy metal, and it accurately converts the sounds of various instruments into MIDI format. Although it can transcribe without knowing which instruments are present, the official team strongly recommends that users provide an instrument list to the model in advance to avoid errors where the model judges instruments inconsistently across 5-second audio segments (flicker). Past models were usually trained only on synthetic audio, making them error-prone when encountering real music. To solve this, the team trained it on 170,000 real recordings. If you are interested in music tech, check out their GitHub or HuggingFace pages to experience the charm of this open-source tool. This is undoubtedly a great boon for music creators.

Google AI Studio Launches Exclusive Custom URL Service

Developers, good news is here. Logan Kilpatrick, head of Google AI Studio, announced that applications deployed on Google AI Studio can now have exclusive, beautiful URLs. The format is approximately “yourname.ai.studio”. Did you know? People often rack their brains trying to think of a memorable domain name. In the past, when deploying free applications, the URLs were always a string of hard-to-remember gibberish. Now, not only are the applications free and deployment free, but custom URLs are also provided for free. This will definitely make everyone’s work look more professional when showcased.

Today we are rolling out pretty URL’s for deployed apps in @GoogleAIStudio, each app can get a “https://t.co/FvJdcNxK0w”, for free!

Free apps, free deploys, and now free pretty URLs! pic.twitter.com/qkxnOyfiTr

— Logan Kilpatrick (@OfficialLoganK) July 10, 2026

Claude Code Desktop Version Adds Built-in Browser to Boost Development Efficiency

Let’s turn our focus to Anthropic. They have brought a very practical update for developers. The desktop version of Claude Code now includes a built-in browser. What does this mean? It means Claude can now directly open official documentation, design drafts, or any website. It can read content and click links, just as naturally as operating on a local development server. These operations are all performed in a sandboxed environment, ensuring safety, and users can also configure whether to retain connection sessions. This improvement really makes the development process much smoother.

Claude Code on desktop now has an in-app browser.

Claude can pull up docs, designs, or any other site. It can read, click through, and interact the same way it does with your local dev servers.

It's sandboxed and configurable: you choose whether sessions persist. pic.twitter.com/Jbc21OP0vJ

— ClaudeDevs (@ClaudeDevs) July 10, 2026

Tech Giant Patent War: Apple Sues OpenAI for Trade Secret Theft

The tech patent war has added another chapter. According to a WIRED report, Apple has formally filed a lawsuit against OpenAI and its hardware executive. Apple accuses them of stealing the iPhone maker’s trade secrets. Specifically, Apple believes OpenAI encouraged hired Apple employees to bring along undisclosed hardware component designs, confidential presentations, secret prototypes, and detailed information on key suppliers. The hardware R&D tug-of-war behind this lawsuit is definitely worth continuous attention. After all, both companies hold significant weight in the tech world, and subsequent developments will certainly impact the entire industry.

How Cognition Builds Trust with Claude Fable 5

The development team for AI software engineer Devin, Cognition, recently shared their views on language models. According to an official Anthropic blog post, Cognition’s VP of Research, Silas Alberti, stated that they have tested almost every generation of Claude models. However, Claude Fable 5 is the first model the team feels confident leaving to operate independently overnight. The thing is, past models tended to lose direction or trigger hard-to-detect errors when handling complex tasks over long periods. But under Cognition’s strict “anti-laziness” testing standards, Claude Fable 5’s performance has significantly jumped over the previous generation. It can maintain clear logic across a messy log record and accurately find root causes. This sense of trust is a key factor in promoting automated development.

Claude Fable 5 Access Extended and Restrictions Relaxed

Following the discussion above, Anthropic has also brought more resources to paid users. An official announcement mentioned that access to Claude Fable 5 will be fully extended for all paid plans. Furthermore, until July 19th, the weekly usage limit for Claude Code has also been significantly increased by 50%. This is definitely a timely rain for teams that need to write a lot of code. Everyone can take this opportunity to fully explore Fable 5’s potential and enjoy more abundant computing resources.

We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.

— Claude (@claudeai) July 12, 2026

ChatGPT Work and Codex Welcome Massive Optimizations and Adjustments

OpenAI hasn’t stopped their momentum either. They have launched a series of important updates for ChatGPT Work and Codex users. First, a quota reset feature has been added, which is now available not just on desktop, but also on web and mobile. Because the system had a short-lived error, the official team generously restored quotas for everyone who pressed the reset button within a specific period, while also taking the opportunity to celebrate the milestone of 7 million active users.

Added a banked reset to 500k users of ChatGPT Work and Codex.

What’s been happening:
- Just released the ability to use a banked reset from web and mobile. Before today you could only do it in the desktop app
- Had an issue where less than 10% of users who used a banked reset…

— Tibo (@thsottiaux) July 12, 2026

Additionally, the team has optimized the inference system. GPT-5.6 Sol subscribers are expected to gain about 10% more usage. However, due to previously increasing the context length limit to 372k, the system consumed more quota than expected. To protect everyone’s usage rights, the official team has temporarily adjusted the limit back to 272k, and will re-evaluate and adjust it in the coming days. At the same time, the team has not only fixed the issue where the multi-agent system consumed excessive quota under high inference load, but also announced they will continue to temporarily cancel the 5-hour usage limit, allowing everyone to enjoy the service more freely. These fine-tuning measures are all for making the overall experience more stable.

Updates for Codex and ChatGPT Work users. No nerfing, only good stuff!

- We have landed inference optimizations and are passing down savings to all the subscriptions for GPT-5.6 Sol. That should result in around 10% more usage on its own.
- We noticed that by changing the… https://t.co/ej7DdJeyBL

— Tibo (@thsottiaux) July 13, 2026

LingBot-VA 2.0 Robot Video Action Model Released

Let’s look at the latest developments in robot control. After six months of development, the team officially released LingBot-VA 2.0. It is a native video action foundation model pre-trained from scratch, specifically designed for a wide range of robot control. Past systems were mostly pieced together from general video generators, but this model demonstrates entirely different capabilities. It mainly brings three key breakthroughs: First, pre-training on native video and actions allows the robot to learn world knowledge, significantly improving generalization capabilities. Second, a semantic visual action tokenizer ensures the accuracy of action prediction. Third, it has forward-looking reasoning capabilities, allowing the robot to think ahead while acting, providing continuous and responsive control. Best of all, it runs in real-time on consumer GPUs, supporting control frequencies up to 150 Hz. This democratization of technology is bound to drive more interesting physical robot applications.

🚀 LingBot-VA 2.0 is here!

After half a year of teamwork, we're thrilled to release LingBot-VA 2.0 — a native video-action foundation model for generalizable robot control.
Unlike prior world-action models that retrofit generic video generators for robot control, LingBot-VA 2.0… pic.twitter.com/mZhBHgPhTw

— Yinghao Xu (@YinghaoXu1) July 9, 2026

LingBot-World-V2 Creates Boundless Interactive Worlds

In addition to physical robot control, the Robbyant team also published the astonishing virtual environment generation model LingBot-World 2.0 (also known as LingBot-World-Infinity). This world model introduces four major upgrades. First is the boundless interactive field of view; through a causal pre-training paradigm, the model maintains stable output quality during continuous interaction. Next is the extremely fast reaction time; the system can instantly drive high-definition video streaming at 720p resolution and 60 frames per second. Most interestingly, it combines diverse interactive elements and an innovative Agentic Harness. In this system, a pilot agent is responsible for planning character behavior, while a director agent is responsible for synthesizing environmental changes. Everyone can even experience its real-time interactive charm on the Reactor or LingGuang platforms.

Q&A

Q1: Why is Apple suing OpenAI? A: According to reports, Apple accuses OpenAI of allegedly stealing its hardware trade secrets. Apple believes OpenAI encouraged hired Apple employees (such as the former hardware executive, etc.) to take along unpublished hardware component designs, confidential engineering presentations, secret prototypes, and detailed information on key suppliers when they left or during interviews.

Q2: What privacy controversy did the Grok Build CLI trigger? A: Cybersecurity investigations found that when the Grok Build CLI runs, it uploads the user’s “entire code repository” (including all tracked files and past git history, even containing unread confidential files like .env) to xAI’s cloud. The most controversial part is that even if the user manually turned off the “improve models” option, this uploading action still proceeds by default.

Q3: What are the features of the open-source music transcription model MuScriptor? Any suggestions for use? A: MuScriptor, developed by Kyutai and Mirelo, was trained on 170,000 real recordings and can accurately transcribe classical, pop, and even heavy metal real music into MIDI format. Although it can transcribe without knowing what’s inside, the official team strongly suggests that users “provide an instrument list in advance” as a condition, to avoid errors where the model judges instruments inconsistently across segments (flicker).

Q4: Why does Devin’s development team, Cognition, rate Claude Fable 5 so highly? A: The Cognition team stated that past models tended to lose direction or produce errors when task execution time was long. But Claude Fable 5 is able to maintain clear logic across messy log records, accurately find root causes, and exhibits an extremely long “task horizon,” becoming the first model that makes the team feel completely confident in allowing it to operate independently overnight for up to 8 hours while continuously advancing progress.

Q5: How is LingBot-VA 2.0 different from past robot control models? A: Past systems were mostly pieced together from general video generators, but LingBot-VA 2.0 is a “native video action foundation model” pre-trained completely from scratch. It not only possesses forward-looking reasoning capabilities (thinking while acting) but also runs in real-time on consumer GPUs, supporting control frequencies up to 150 Hz.