Daily AI News | Alibaba Cloud revenue grows 45% as AI products post a twelfth triple-digit quarter
# Daily AI News | Alibaba Cloud revenue grows 45% as AI products post a twelfth triple-digit quarter
## Lead Story
### Alibaba Cloud revenue grows 45% as AI products post a twelfth triple-digit quarter
Alibaba reported that external commercial revenue at Alibaba Cloud rose 45% year over year and adjusted EBITA climbed 133%. The company said AI-related product revenue recorded triple-digit growth for a twelfth consecutive quarter, while AI cloud and computing-services revenue reached RMB 48.437 billion.
**AI take:** The figures show model and compute investment becoming material revenue, though durability depends on sustained enterprise usage.
## More News
### Binance launches Agent OS for AI trading under user-defined controls
Binance launched Agent OS and an MCP Server that allow AI agents to analyze markets and execute trades under permissions and controls set by users. The system can work with tools including ChatGPT, Claude Code, and Cursor.
**AI take:** Convenience now reaches real assets, making permissions, limits, and human confirmation essential safeguards.
### Research shows encrypted malicious instructions can make Grok exfiltrate data
Ars Technica reported that an attack called Cryptographic Context Injection can hide malicious instructions in encrypted text, bypass Grok safeguards, and induce the model to exfiltrate user data. The result shows that encrypted content can still carry prompt injection.
**AI take:** Model filtering alone is insufficient when private data or tools are connected; least privilege and output isolation remain necessary.
### Tencent Cloud launches TDSQL Nexa to unify enterprise data for agents
Tencent Cloud launched TDSQL Nexa, combining transactions, full-text and vector search, AI compute, and analytics in one data platform. It also provides business semantics, separate agent identities, row- and column-level permissions, and auditing; Tencent says several related capabilities are already in production.
**AI take:** Enterprise-agent bottlenecks are shifting from models to data permissions and semantics, with customer deployments the next test.
### Meta launches a macOS Meta AI app with cross-app dictation
Meta launched a Meta AI application for macOS. The company says its dictation feature works across applications, allowing users to speak into whichever app they are using.
**AI take:** A desktop surface reduces friction, but accuracy and privacy controls will determine whether it replaces established dictation tools.
### Meta opens AI game-creation app Pocket to US users
Meta opened its experimental AI application Pocket to users across the United States after quietly testing it in Brazil. Pocket lets users create and share interactive games with natural-language tools.
**AI take:** It brings low-code creation to consumers, with output quality, moderation, and creator retention as the next tests.
### Ramp launches Router to switch among AI models through one API
Ramp launched an AI model-routing service called Router. Users and companies can access and switch among different large language models through a single API.
**AI take:** A unified endpoint can cut integration work, but customers still need to inspect routing logic, pricing, and data handling.
### Liquid AI releases LFM2.5-DSpark optimization claiming up to 3.2× faster inference
Liquid AI published an LFM2.5-DSpark inference optimization on Hugging Face and said its tests achieved speedups of up to 3.2 times. The figure is a publisher benchmark, and real gains depend on hardware and workload.
**AI take:** If reproducible, the improvement can cut local-serving latency and cost; independent benchmarks are the next evidence.
### Google lets readers mark publishers as preferred sources amid AI traffic losses
Google added a button that lets readers mark publishers as preferred sources across Search, Discover, and Google News. The feature gives publishers a new audience signal as AI search sends fewer referral clicks to the web.
**AI take:** It offers publishers a distribution lever, but results still depend on reader adoption and Google’s ranking weight.
### Baidu opens advanced agent capabilities for free, moving from answers to deliverables
Baidu opened advanced agent capabilities for free, extending the product from search and question answering toward completing and delivering tasks. The release emphasizes multi-step execution rather than answers alone.
**AI take:** Free access lowers the trial barrier, while success rates and controllability on complex tasks will determine real value.
### Doubao large model is integrated into Tesla vehicles in China
LeiFeng reported that the Doubao large model has been integrated into Tesla vehicle systems in China. The move brings ByteDance model capabilities into an in-car setting, while supported vehicles, functions, and rollout scope still need clarification.
**AI take:** The in-car surface expands Doubao’s reach, but value depends on vehicle coverage, reliability, and permission boundaries.
### Simon Willison tests smolvm for sandboxing untrusted Python and JavaScript
Simon Willison tested smolvm as a sandbox for untrusted Python and JavaScript, focusing on memory, CPU, network, and file-access restrictions. Because Claude Code for web lacked KVM, the tests were moved to a GitHub Actions runner with virtualization support.
**AI take:** It offers a reviewable path for agent-generated code, but production use still depends on KVM and strict resource policies.
## Rumors
### Anthropic reportedly plans to let enterprises retain model data in their own clouds
Bloomberg and Reuters reported that Anthropic plans to keep a 30-day retention requirement for some frontier models while allowing enterprise customers to hold that data in their own cloud infrastructure. The safety system is reportedly due later this year, but final terms have not been announced.
**AI take:** If implemented, it could reduce data-location concerns while leaving enterprises responsible for 30-day retention and auditing.
