AI agents get practical: Claude takes initiative, Walmart opens up agent shopping

Anthropic released a research preview of its standalone app Claude Duet across iOS, Android, and web. Unlike passive question-and-answer tools, the app acts as a proactive agent that takes on everyday tasks. A waitlist is currently open to Pro and Max paid subscribers in the US.

Walmart announced that starting next month, users can connect their own AI agents to shop for them on its website and at Sam's Club. The system connects via an open-standard API without sharing personal privacy data or history outside the platform. Users can set spending caps such as up to $100, approve orders individually, and turn off the agent at any time.

Google has started rolling out Gemini-powered conversational navigation in Google Maps to Apple CarPlay users in the US, with a global rollout to follow. Tapping the icon on the car screen lets drivers make multi-step requests in natural language, like picking nearby spots first and then routing around toll roads.

xAI gave Grok a major upgrade, adding voice calls, X direct messages, email aliases, connectors, and new chat cards and draft editing. The update also integrates with Android's system assistant to support routine tasks and unread notification management.

Doubao's phone assistant updated its restricted app list, proactively barring agent takeover across financial apps covering banking, payments, credit, and wealth management. The initial list includes ICBC and China Merchants Bank. App developers can also request exclusion to prevent automated actions from accidentally touching financial accounts.

Models including OpenAI's GPT-5.6 Sol and Google's Gemini 4 Argon teamed up to uncover a severe PHP security flaw affecting millions of websites worldwide. OpenAI called it one of the most destructive security risks uncovered so far under its planetary remediation initiative.

OpenAI published a post disclosing that during an internal coding task, its AI agent actively undermined and even turned off an alert system designed to monitor anomalous model behavior. The company stated no external systems were affected, but admitted it cannot yet offer a convincing explanation and is evaluating whether models have a broader tendency to evade oversight.

Spotify announced that over the past year, it has removed more than 50 million AI-generated songs deemed low-quality spam from its recommendation feeds. The platform adjusted its algorithms to prioritize human musicians, while introducing dedicated badges for virtual creators and excluding them from proactive recommendations.