The iPhone Duo should have been an iPad

What It Covers

The Vergecast hosts David Pierce and Nilay Patel debate the iPhone Duo's 7/10 review score, argue the $2,000 foldable functions better as an iPad than an iPhone, and cover two major tech crises: a PHP security vulnerability discovered via AI that threatened 70% of the web, and Meta's terms-of-service update attempting to scrape publisher websites for AI training.

Key Questions Answered

• •iPhone Duo app optimization: The Duo operates on three distinct app tiers — unoptimized iPhone apps floating in black space, iOS 27-updated apps with black side bars, and full iOS 27.1 Duo-specific layouts. Only the third tier delivers a functional experience, and most major apps, including Instagram and Gmail, remain broken or partially optimized weeks after launch. Buyers should expect a 6-to-12-month window before app support meaningfully improves across the ecosystem.
• •Foldable device positioning: The iPhone Duo's outer cover screen is too small for practical daily use, while the inner display delivers a compelling tablet experience. Reviewers including John Gruber and Marques Brownlee independently concluded the device functions as a flawed iPhone but a superior iPad. Apple's decision to brand and sell it as an iPhone alongside standard models creates mismatched expectations for consumers who are not self-selecting early adopters.
• •PHP zero-day AI discovery: Security firm Wiz Research used the prompt "find bad vulnerabilities in PHP" to discover critical memory-overflow exploits in PHP, which underlies approximately 70% of websites globally. A second firm, ChainGuard, independently replicated the discovery and released patches before the coordinated disclosure embargo, triggering a worldwide scramble. Government websites across the Netherlands and Finland went offline, Wikipedia experienced outages, and Cloudflare logged millions of exploit attempts within days.
• •AI-accelerated cyberattack economics: Exploiting the PHP vulnerability previously required large teams manually targeting specific server states — a practically infeasible attack at scale. AI agents can now automate this process continuously and cheaply. The shift means offensive security operations that once required significant human resources are now accessible to any attacker willing to type a single sentence prompt, fundamentally changing the cost structure of large-scale web exploitation.
• •Agentic commerce friction breakdown: Meta's Muse AI agent has surpassed 10 million users, with 25% of active users engaging daily and a quarter of shopping users averaging purchases over $200. However, agents performing tasks like repeatedly holding shopping cart items on platforms like The RealReal — locking products for hours without purchasing — replicate Ticketmaster bot behavior at scale. Platforms face a binary choice: block all agents entirely or negotiate opaque commercial access deals.
• •Meta terms-of-service AI training overreach: Meta quietly updated business account terms to authorize scraping any website, URL, or domain associated with a Meta business account to train AI models. The language explicitly permitted bots, crawlers, and spiders to collect and use content from linked publisher sites. After public backlash and legal analysis confirmed the scope, Meta revised the terms and claimed robots.txt compliance would be honored — but the original contractual language remained more legally binding than the subsequent social media clarifications.
• •Arc browser shutdown and browser-layer AI failure: The Browser Company's Arc browser will officially shut down in December, following a pivot to AI-focused browser Dia after acquisition by Atlassian. Arc's vertical tabs and sidebar design influenced Chrome, Zen, and OpenAI's Atlas browser. The broader lesson: the strategy of inserting AI at the browser layer to access user data has been outpaced by OS-level AI agents and cloud Linux environments that bypass the browser entirely, making browser-as-AI-platform a losing competitive position.