Apple Sues Sam; Meta's Muse Meltdown; Grok Grabs Your Keys

Today's AI Outlook: 🌥️

Apple Tries To Stall OpenAI’s Hardware Dreams

Apple has sued OpenAI, hardware chief Tang Tan, former Apple engineer Chang Liu and OpenAI’s io devices unit, alleging the company’s push into consumer hardware benefited from stolen trade secrets. Apple wants the court to force changes to OpenAI’s unreleased, Jony Ive-designed device, which is reportedly expected in 2027.

The complaint turns an already tense talent war into a potential threat to OpenAI’s hardware timeline. Apple says more than 400 former employees now work at OpenAI and claims some recruits arrived with confidential information, internal files and knowledge of unreleased products.

Why it matters

OpenAI is preparing to compete with the company that defined modern consumer hardware. A long court battle could delay its device, force a redesign or surround the launch with questions about how its technology was developed.

The Deets

• Apple says it warned OpenAI about suspected theft in February and received no response.
• Tang Tan spent 24 years at Apple before joining OpenAI.
• Apple alleges candidates were encouraged to bring “actual parts” into hardware interviews.
• Liu is accused of retaining access to internal cloud files after leaving Apple.
• OpenAI said it has “no interest in other companies’ trade secrets.”
• The device is being developed with Ive, Apple’s former design chief.

Key takeaway

Apple may struggle to stop OpenAI’s hardware ambitions, but it can make the road to launch slower, costlier and legally messy.

🧩 Jargon Buster - Trade secret: Confidential business information that provides a competitive advantage, such as designs, manufacturing methods or unreleased product plans.

♟️ Power Plays

Meta’s Superintelligence Debut Lasted Three Days

Meta removed Muse Image’s tag-to-remix feature within 72 hours of its launch after users and the actors’ union SAG-AFTRA criticized the product. The feature allowed Instagram users to apply AI-generated edits to public photos, including images posted by other people.

Meta enabled the capability by default for public accounts, creating immediate concerns about consent, impersonation and image manipulation. The company acknowledged that it had “missed the mark” and pulled the feature.

Why it matters

Meta’s first major product from its Superintelligence Labs became a case study in how quickly an AI feature can collapse when consent and product safeguards arrive after launch.

The Deets

• Muse Image launched on July 8.
• Users could remix other people’s public Instagram photos with one tap.
• Public accounts were included by default.
• SAG-AFTRA condemned the feature over potential misuse.
• Meta withdrew the capability after widespread criticism.
• The feature was tied to Meta’s broader push toward a “personal superintelligence.”

Key takeaway

AI product teams can ship impressive capabilities quickly, but platforms still need clear permission systems before letting users alter someone else’s identity or likeness.

🧩 Jargon Buster - Opt-out: A system that activates a feature automatically and requires users to disable it themselves.

🛠️ Tools & Products

Grok CLI Packed More Than The Prompt

A researcher examining xAI’s official Grok command-line tool found that it created before-and-after archives of a project and uploaded them to an xAI cloud bucket through a separate pipeline. The transfer reportedly occurred even when the model was asked to respond with a single word.

During a test on an isolated repository, the upload reportedly captured the project’s code, a Claude Code configuration from another folder and a live API key that had not been intentionally shared with Grok. xAI later disabled the upload mechanism from its servers without publicly defending the practice.

Why it matters

Coding agents often need broad access to local files, terminals and credentials. Undisclosed collection creates serious risks for developers working with private repositories, customer data or production keys.

The Deets

• The tool created compressed tar.gz archives of the project.
• It captured files before and after the agent performed its task.
• Uploading occurred through a channel separate from the model response.
• The archive reportedly included information outside the test repository.
• A functioning API key was included in the transfer.
• xAI disabled the mechanism after the findings circulated.

Key takeaway

Developers should treat local AI agents like powerful software integrations and verify exactly which folders, credentials and network connections each tool can access.

🧩 Jargon Buster - CLI: A command-line interface, which lets users operate software by typing commands into a terminal instead of clicking through a graphical app.

💰 Funding & Startups

Manus Gets Pulled Back Into China’s Orbit

Tencent is reportedly leading a consortium seeking to acquire Manus, the Chinese AI agent startup, at the same $2B valuation Meta paid for it in December 2025. The move follows an order from Beijing reversing the U.S. deal amid concerns that strategically important AI companies were being transferred overseas.

Tencent would reportedly become the largest shareholder while remaining a minority owner. Manus would continue operating independently from Singapore, with a possible Hong Kong listing ahead. Its founders, however, are reportedly barred from leaving China.

Why it matters

AI startups increasingly sit inside a geopolitical contest over intellectual property, talent and national technological capacity. Government intervention can now determine who is allowed to acquire a company, even after a deal has closed.

The Deets

• Meta reportedly bought Manus for $2B in December 2025.
• Beijing ordered the transaction reversed.
• Tencent is leading the group seeking to buy the company back.
• Manus reportedly generates $500M in annual recurring revenue.
• The company could pursue a Hong Kong listing.
• Its Singapore operations are expected to remain independent.
• Manus’ founders reportedly cannot leave China.

Key takeaway

For strategically important AI companies, investors and founders may no longer control the final destination of an acquisition.

🧩 Jargon Buster - Annual recurring revenue: The yearly value of predictable subscription or contract revenue a company expects to receive.

🧪 Research & Models

AI Forecasters Draft A Plan For The 2030s

The AI Futures Project, the nonprofit behind AI 2027, released AI 2040, a collection of five scenarios exploring how the global AI race could develop over the next decade. The paths include continued competition, unilateral delays, geopolitical confrontation, international cooperation and a worldwide research halt.

Its recommended scenario, called Plan A, proposes that the U.S. and China pause frontier-model training in 2029, expand international chip tracking and require leading laboratories to make more of their research visible.

Why it matters

Forecasts cannot predict the future, but they can expose decisions governments and AI companies may face as models become more capable. The project focuses attention on transparency, verification and coordination before a crisis forces those systems into place.

The Deets

• The five scenarios begin from a shared decision point in 2029.
• Plan A proposes a coordinated U.S.-China pause on frontier training.
• The proposal includes expanded global tracking of advanced chips.
• Leading laboratories would face greater research transparency.
• Co-author Daniel Kokotajlo expects companies to resist those requirements.
• The earlier AI 2027 scenario reached Vice President JD Vance, who expressed concern about its conclusions.

Key takeaway

AI 2040 offers policymakers a structured way to test how competition, regulation and international cooperation could interact before the highest-stakes decisions arrive.

🧩 Jargon Buster - Frontier model: One of the most advanced AI systems available, typically requiring substantial computing power, data and research investment.

⚡ Quick Hits

• OpenAI leadership: Fidji Simo, OpenAI’s CEO of AGI deployment, is moving into a part-time advisory role because of a chronic illness.
• Safety departure: OpenAI Head of Safety Johannes Heidecke is reportedly leaving the company, following Chief Futurist Joshua Achiam’s departure.
• Math milestone: OpenAI said GPT-5.6 Sol Ultra created 64 subagents and generated a proof for a 50-year-old unsolved math problem in under an hour.
• Security surge: Microsoft warned that AI-assisted vulnerability discovery could make Patch Tuesday considerably busier as automated systems uncover more software flaws.
• Chip ambitions: Apple is reportedly repurposing work from its discontinued autonomous-vehicle program to accelerate development of M7 Ultra AI chips supporting as much as 1.5TB of RAM.
• Memory money: SK Hynix raised $26.5B in what AI Secret described as the largest foreign IPO in U.S. history, driven by demand for AI memory.
• Owning the stack: Hugging Face said more companies are exploring open-source AI systems they can operate themselves instead of continuously renting access to frontier models.
• Synthetic feeds: Pangram reported that long-form posts on LinkedIn and X contain increasingly high levels of AI-generated writing.

🧰 Tools Of The Day

ChatGPT Work + Codex: OpenAI’s desktop workflow can turn a written product brief into a working website. Users can create a project folder, have Work coordinate research subagents, then send the resulting plan to Codex for development and visual fixes. See the guide in The Rundown AI.

Strands Agents: AWS’ open-source agent SDK includes context management, execution limits, hooks and observability for building production agents across different models and environments.

Framer Agents: Framer allows coding agents such as Claude Code, Cursor, Codex and Gemini CLI to edit site layouts, components, responsive settings, CMS content and SEO configurations while keeping changes visible and reviewable.

Runway Dev: A unified API providing access to Runway’s video models, real-time avatars and third-party image and audio generators.

Adapt: A Slack-native AI coworker designed to connect with business systems and complete workplace tasks inside Slack.

Today's Sources: The Internet, The Rundown AI, AI Secret