Codex Gets a Dedicated Windows Sandbox with Filesystem Boundaries and Firewall Rules
OpenAI has shipped a dedicated Windows sandbox environment for Codex, its AI software engineering agent. The sandbox gives Codex a contained execution environment on Windows with configurable filesystem permission boundaries, so teams can specify exactly which directories the agent can read, write, or execute within. Firewall rules are also configurable at the sandbox level, meaning security-conscious teams can block outbound calls to arbitrary endpoints during code generation and testing.
This is directly relevant to enterprise adopters who have held back from using Codex on sensitive codebases precisely because the agent's file access was hard to scope. The new sandbox makes Codex usable in environments where IT policy requires a clear boundary between agent activity and production systems. OpenAI announced this on its own channels before mainstream technology press coverage, which fits the pattern of developer-facing releases being pushed direct rather than via PR.
Claude for Small Business Launches with 15 Workflows, Major App Connectors, and a 10-City Tour
Anthropic has launched Claude for Small Business today, a purpose-built offering aimed at owner-operators and small teams rather than enterprise IT departments. The package ships with 15 pre-built workflow templates covering common small business tasks (proposal drafting, invoice follow-up, customer onboarding, social content, financial summaries) and 15 skills (discrete capabilities that can be layered into those workflows). Connectors cover the tools most small businesses already use: QuickBooks, Canva, Google Workspace, Microsoft 365, Slack, DocuSign, HubSpot, and PayPal.
Crucially, Anthropic is not charging extra licensing fees for the integrations. The connectors are included in the existing subscription tier, which removes the calculation small business owners would otherwise have to make about whether the added cost justifies the setup effort. The go-to-market motion includes a 10-city training tour starting today, which suggests Anthropic is investing in in-person adoption rather than assuming small business owners will self-serve through documentation alone.
The timing is worth noting: this lands two days after OpenAI announced a $4 billion enterprise deployment arm targeting large organisations. Anthropic appears to be explicitly flanking that move by going after the market that OpenAI's new subsidiary is least likely to serve well. For a business owner who currently uses a mix of separate SaaS tools, the pitch of a single AI layer that connects them without requiring a system integrator is genuinely compelling.
Magic Pointer Makes the Cursor a Gemini-Aware Input Signal in Chrome
Google has announced Magic Pointer, a new capability for Gemini in Chrome that treats cursor movement as a contextual signal rather than just a navigation tool. As you hover over elements on a webpage, Gemini reads the intent implied by your cursor position alongside the page content and any voice or text prompt you have active. The result is that you can point at a passage, a table cell, or a product image and ask a question without having to copy, paste, or explicitly describe what you are referring to.
The system processes hover signals locally where possible, with Gemini calls made only when a question is posed or a persistent action is requested. This sidesteps the obvious privacy concern of continuous cursor telemetry being sent to a server. For everyday browsing the difference may be subtle, but for research workflows involving dense documents, product pages, or data tables, the ability to point and ask without switching context is a meaningful reduction in friction. Magic Pointer is rolling out via Gemini in Chrome and will be available to Gemini Advanced subscribers from today.
Google's Android Show yesterday previewed Gemini Intelligence (adaptive on-device reasoning), Rambler (a spatial audio journalling feature), and Googlebook (an AI-curated reading companion). Full details are expected at Google I/O on 19 May.
WhatsApp Adds Incognito Chat for Meta AI with Private Processing and Text-Only Mode
Meta has added Incognito Chat to WhatsApp's Meta AI integration. When a user activates it, their conversation with Meta AI is processed using Private Processing: a confidential computing approach that prevents Meta's servers from seeing the content of the exchange in plaintext. Incognito Chat is text-only for now, which avoids the harder problem of processing images or voice in a privacy-preserving way, and conversations are not stored or used for model training.
A Side Chat mode is also in the pipeline, which will let users hold an AI conversation that runs alongside a human chat thread without the AI's contributions appearing in the shared conversation. Both features address a real hesitation among WhatsApp users who want AI assistance but are uncomfortable with the idea that their queries are visible to a platform that also holds their personal communications. Incognito Chat is rolling out globally from today; Side Chat has no confirmed date.
Qwen Powers Conversational Shopping Across 4 Billion Taobao and Tmall Listings
Alibaba has integrated Qwen directly into Taobao and Tmall, giving shoppers a conversational interface across the entire combined catalogue of roughly 4 billion product listings. Rather than searching by keyword and filtering by category, users can describe what they want in natural language and Qwen surfaces products, compares specifications, flags relevant reviews, and can place an order on their behalf once they confirm. The integration also includes Wukong, Alibaba's enterprise agent for business sellers, which uses Qwen to generate product descriptions, manage promotional pricing, and draft responses to customer queries at scale.
The scale here is genuinely difficult to contextualise for a Western audience. Taobao and Tmall together serve well over a billion active users. Even a modest adoption rate of the conversational interface represents a larger AI e-commerce deployment than anything operating in Europe or North America. If the user experience is good, this could set the expectation for what online retail feels like globally within 18 months.
Edge Retires Copilot Mode and Folds Its Features into the Default Browser Experience
Microsoft is retiring the dedicated Copilot Mode in Edge and absorbing its capabilities into the standard browser. The features that lived behind Copilot Mode (multi-tab reasoning, Voice plus Vision, Study and Learn, and Journeys) are moving into Edge's default interface, available to all users without switching modes. The change removes the cognitive overhead of deciding when to activate Copilot Mode and positions AI assistance as a background capability rather than an explicit tool you switch to.
This is the logical end of a mode that always sat awkwardly between the browser and the AI assistant. Folding it into the default experience means the features get used more (people do not have to remember to switch on a mode) and removes the implicit message that AI assistance is an optional extra rather than a core part of how the browser works. Worth noting that Workday Sana Agent also arrived in Microsoft 365 Copilot today, adding Workday HR workflows directly into the Copilot chat interface without requiring users to leave M365.
xAI is reportedly in discussions with Apollo Global Management, Morgan Stanley, and Valor Equity Partners about a finance-focused push for Grok. No product announcement yet; this appears to be early commercial conversations rather than a launch.
Mistral is developing a cybersecurity-focused model specifically for banks and financial institutions. No release date has been confirmed but the project is understood to be well advanced.
Reports are circulating that DeepSeek is close to closing a funding round of $3 to $4 billion at a valuation of around $50 billion. Nothing confirmed from DeepSeek directly; treat this as market intelligence rather than fact until an announcement.
Perplexity published a security changelog dated 13 May covering several improvements to data handling and session management. Routine transparency update rather than a response to an incident.
Industry Themes
The most strategically interesting story today is Anthropic's small business launch, and not just because of the product. The timing, two days after OpenAI announced its enterprise deployment arm, signals a deliberate segmentation strategy. OpenAI is building infrastructure to serve the 5,000-seat enterprise; Anthropic is building workflows for the 5-person shop. Both can be right simultaneously, and the total addressable market for the latter is considerably larger by number of organisations, even if the revenue per account is lower.
The Qwen deployment across Taobao and Tmall deserves more Western attention than it is likely to receive. When a conversational AI interface becomes the primary way a billion-plus people interact with the world's largest e-commerce catalogue, it sets expectations that will ripple outward. The question for every retailer outside China is not whether this is coming to their market, but when and from whom.
Google's Magic Pointer and Microsoft's Edge changes both reflect the same insight: AI assistance works best when it is ambient rather than modal. Switching into a special AI mode adds friction that compounds into non-adoption. Treating the AI layer as part of the default experience, whether that is a browser or a cursor, is where the usage will actually accumulate. The labs that crack ambient integration will have a usage moat that pure model quality cannot overcome.