Last week, a ransomware attack was widely reported as the world's first fully autonomous AI-driven cybercrime. However, a closer look by TechCrunch reveals a more nuanced picture. While an AI agent did carry out the technical execution steps, a human was still responsible for choosing the target, setting up the infrastructure, and providing stolen login credentials. In other words, the attack was far from fully autonomous — a human criminal remained in control of the key decisions. This case is a clear reminder of how dramatic headlines about AI can easily outrun the actual facts.

Google has announced a significant expansion of its Managed Agents feature in the Gemini API. Two of the most important additions are remote MCP (Model Context Protocol) support and background task execution. Remote MCP allows agents to connect with external data sources and tools through a standardized protocol, which means developers no longer need to build custom connectors from scratch. Background task execution lets agents handle long-running processes asynchronously, without waiting for user input at each step. These updates suggest that the real battleground in the AI agent race is shifting from raw model performance to infrastructure and ecosystem depth.

In an interview with TechCrunch, Vercel CEO Guillermo Rauch made a strong case for separating the model layer from the agent layer in AI system design. He explained that once teams start optimizing for production environments, they inevitably begin weighing price against performance. By keeping the agent architecture independent from any single model provider, developers gain the flexibility to choose or switch models based on the specific needs of each task. This approach also makes it easier to avoid vendor lock-in and keep costs under control — a practical and timely concern for any team building large-scale AI applications.

Anthropic has announced that Claude Cowork, its AI collaboration tool, is expanding to mobile and web platforms. Previously available only as a desktop application for macOS and Windows, the service will roll out first to Max plan subscribers, with access for other plans expected to follow within a few weeks. This move is more than a simple platform expansion. As competition with ChatGPT and Gemini intensifies, being available anytime and on any device has become a critical factor in adoption decisions for both enterprise and individual users. The launch signals that Anthropic is steadily evolving from a model provider into a broader AI platform company.