Meta’s Muse agent hits No. 2, but there’s a big privacy catch
Unlike standard chatbots that just answer questions, Muse acts on your behalf—booking travel, paying bills, and managing your calendar by directly accessing your accounts.
- Isolated cloud execution: To run web tasks safely, Meta spins up an isolated virtual machine (VM) with a sandboxed browser for every user session. This hides your raw passwords and credit card details from the core AI model.
- Secondary security layer: A dedicated safety guardrail sits between the agent and the web, reviewing and approving every transaction before it goes through.
- The current security gap: Even though your credentials are masked from the AI, Meta engineers can still view data inside the active VM session. Full end-to-end encryption for these virtual environments is not expected until later this year.
This launch signals a clear shift from conversational search to actual task delegation. While the productivity benefits of letting software handle your daily errands are obvious, it requires handing over an unprecedented level of account access to a single tech company.
If you decide to test Muse, stick to low-risk tasks like schedule management first, and hold off on linking sensitive financial accounts until end-to-end encrypted VM sessions roll out.
Prime Video wants you to shop while you stream
Amazon is aggressively blurring the line between streaming and checkout, but turning passive TV viewing into a continuous digital storefront risks sacrificing the viewing experience for immediate retail conversion.
- 8,000+ shappable titles: The “Shop the Show” catalog expanded from 1,300 titles across movies, series, and live sports broadcasts.
- Lens visual matching: Amazon Lens powers a new “Shop the Scene” feature that scans outfits or decor on screen and instantly pulls up matching items on your phone.
- Real-time app sync: Opening the Amazon mobile app while streaming automatically syncs your phone feed to the exact scene playing on your TV.
This integration leverages Amazon’s retail backbone to bypass traditional product search, capturing impulse buys the exact second a viewer sees something they like.
Watch for streaming rivals to scramble for similar visual-search partnerships as shoppable media becomes standard across ad-supported tiers.
Jensen Huang bets on AI cybersecurity as the next wave
Nvidia is pushing cybersecurity as AI’s next massive market, but making tools that write code faster ultimately forces companies to buy AI defense software just to fix the bugs their own AI created.
- The 27-second breach: Attacker breakout times have dropped to under half a minute, making human security teams completely useless against automated attacks.
- Creation fuels demand: Huang admitted that faster AI coding squeezes the time between a bug appearing and getting exploited, creating instant demand for auto-patching tools.
- Heavyweight alliances: Nvidia is backing this up with real tech, partnering deeply with CrowdStrike, Cisco, and Palantir using its Nemotron models.
As software development moves at lightspeed, old-school security checks and manual responses just cannot keep up with machine-speed threats.
If you build or manage software, start using continuous AI security scanning now—relying on manual checks will leave you wide open.
DeepSeek Launches V4.1-Flash Model
DeepSeek has released V4.1 Flash, a smaller model designed for greater capability, faster inference, higher throughput and scaling to larger models.
- V4.1 Flash uses a new Causal-Encoder-Decoder architecture with a 552 billion parameter framework, but its MoE design activates only 8 billion parameters for inputs and 16 billion for outputs.
- DeepSeek says the model outperforms its V4 Pro on coding, cybersecurity, and autonomous agent tasks.
- On Terminal-Bench 2.1, V4.1 Flash scored 90.6, ahead of GPT-5.6 Sol at 88.8, Kimi K3 at 88.3, and V4 Pro at 87.9.
- The model also scored 88.1 on Cybergym, beating V4 Pro, Kimi K3, and GPT-5.6 Sol on the cybersecurity benchmark.
- DeepSeek reduced the model’s key-value (KV) cache from 3,514 to 890 bytes per token, significantly lowering inference costs for AI agents.
DeepSeek makes the model files available on Hugging Face under the open MIT license, and also through its API at the same prices as V4-Flash. The launch highlights China’s growing focus on making advanced AI faster and cheaper to run.
Meta Launches Muse as Its New Personal AI Agent
Meta has launched Muse, a personal AI agent designed to help with everyday tasks and projects, like online shopping, sending emails, and planning a trip. The company sees it as an early step toward its broader vision of “personal superintelligence.”
- Muse, powered by the company’s in-house model Muse Spark, runs through a chat interface and lets users name their agent, create an avatar, and customize its communication style.
- The AI agent works by connecting to the user’s apps and services that are part of everyday workflows, like email, calendars, payments, health apps, shopping, music, and more.
- Each agent operates on a dedicated virtual machine in Meta’s cloud with a built-in browser that users can see.
- Muse can continue working in the background even after users close the app, returning to users if something changes or if it requires approval for something.
- Muse is initially available in the US on iOS, Android, and the web, with support for Meta’s AI glasses planned. Meta is also offering a free tier alongside $20 and $100 monthly plans.
Muse arrives as AI agents are gaining more access to users’ accounts and digital lives. The bigger question for Meta is whether consumers will trust an agent with that level of access while finding it useful enough to justify the risks.
Transform Your Drawings into Polished Images with ChatGPT’s New Sketch Feature
The new Sketch feature lets you draw right in ChatGPT and use it as a visual guide.
Step 1: Go to ChatGPT and sign in.
Step 2: Activate the Sketch feature by typing @Sketch into the chat box, and a window will pop up where you can make your drawing (you can add text and shapes).
Step 3: Once you’re done drawing, enter a prompt describing the style and any other details you want.
It can be something simple like: “Make a realistic version of this drawing.”
If you need a helping hand, you can check the built-in image templates, such as anime, sticker, icon, etc. Just select Images on the left, choose Templates, and pick the template you like.
Step 4: Once ChatGPT generates the image, you can download it or click “Edit” to make changes.
You can resize, remove the background, and also leave comments directly on parts of the image that you want changed.
Gemini now has a Windows app
Following the macOS launch in April, Google is bringing the Gemini desktop app to Windows. With the app, users can press Alt + Space to pull up Gemini whenever they need it.
Universal Music Group and ElevenLabs partner on AI music
Universal Music Group is launching an AI music platform with ElevenLabs that will let users create remixes, mashups and new versions of licensed tracks. Artists can choose whether to participate, with the partnership also planned to support new AI products and fan experiences.
OpenAI math controversy deepens after another mathematician speaks out
Mathematician Andreas Thom has accused OpenAI of being dishonest and failing to disclose whether researchers’ private work or ChatGPT interactions contributed to its recent mathematical breakthroughs. His comments add to growing concerns among mathematicians that OpenAI could use unpublished research to compete with the same researchers without consent or proper credit.
Apple CEO says the iPhone is still the best AI device
Apple CEO John Ternus says the iPhone is already the ideal “intelligent personal hub” for AI, combining on device processing, cloud access, cameras, microphones and seamless integration with other devices. He also emphasized privacy as a key differentiator, saying Apple Intelligence runs on device whenever possible to keep users’ data under their control.