SAP announced it will complete the acquisition of master data management provider Reltio in the second or third quarter of 2026 (terms undisclosed), integrating its cloud-native, AI-first MDM and entity resolution capabilities into SAP Business Data Cloud. SAP says this will help customers unify SAP and non-SAP data, generate high-fidelity 'golden records,' provide traceable data context for agent-based AI workflows, and support integrations like the Model Context Protocol (MCP); Reltio will retain its independent sales model.

Fortune reported that Anthropic exposed unreleased model materials and internal documents (including images, PDFs, and event materials) due to misconfigured content management and data storage settings, making them publicly accessible. After the incident came to light, Anthropic confirmed it is testing a new model codenamed 'Mythos,' which offers significant capability improvements over existing products and has already been made available for testing to select early customers. The company stated the leak did not involve customer data or core systems and that remediation and access controls have since been completed.

NPR reported that U.S. District Judge Rita F. Lin of the Northern District of California issued a preliminary injunction temporarily halting the Trump administration's directive requiring federal agencies to stop using Anthropic and designating it a 'supply chain risk.' The judge noted such designations are typically reserved for foreign intelligence or terrorist organizations, not domestic U.S. companies; the record suggests the government’s action may be retaliation for Anthropic’s public criticism of military use of Claude, potentially violating the First Amendment. The injunction remains in effect during litigation, and the government may appeal.

InfoQ reported that OpenAI has expanded its Responses API to serve as a foundation for autonomous agents: the API includes a built-in 'execute-feedback' loop, enabling models to propose actions, execute them, and iteratively refine based on outcomes within a controlled environment until tasks are completed. A new Shell tool supports Unix commands and programs in Go, Java, and Node, along with a containerized workspace and centralized policy layer that keeps credentials invisible to the model and enforces controlled network access; long-running task state is maintained via context compression.

Kunlun Tech unveiled three models—Matrix-Game 3.0, SkyReels V4, and Mureka V9—at the Zhongguancun Forum, targeting AI-powered gaming worlds, video generation, and music generation respectively. Materials indicate Matrix-Game 3.0 supports minute-long continuous interaction at 720p, focusing on long-horizon consistency and real-time performance; SkyReels V4 uses a native audiovisual dual-stream MMDiT architecture and leads in audio-video generation on the Artificial Analysis benchmark; Mureka V9 introduces MusiCoT to enhance controllability in music creation. The company also introduced its '3+1' strategy, using a super agent to orchestrate the three scenario-specific models.

36Kr reported that the draft 'Automotive Software Quality and Defect Management Specification (for public comment)' is now open for consultation, described as the world's first automotive-grade standard integrating functional safety, SOTIF, cybersecurity, and AI safety. The draft mandates quality-by-design and critical process reviews throughout the software development lifecycle, with specific requirements for AI software including value alignment controls, model robustness assurance, and dataset traceability. For high-risk defects, levels 4 and 5 require mandatory recalls, and automakers must report recall plans within five working days.

Turing Corporation announced real-time autonomous driving control on public roads using a Vision-Language-Action (VLA) model: its proprietary ~2 billion-parameter model performs 10Hz real-time inference and vehicle control on an onboard computer, verified stable in real-world environments. The company also released on Hugging Face the RACER-Mini dataset for enhancing causal reasoning and the pre-trained DriveTiTok image tokenizer, which compresses driving scenes by approximately 1/100. This work is part of the GENIAC project supported by METI and NEDO.

In the March Gemini Drop update, Google launched 'Import Memory/Chat History,' allowing users to migrate preferences and past conversations from services like ChatGPT and Claude to reduce switching costs: users can paste exported content or upload ZIP chat history files up to 5GB, with Gemini automatically organizing the data; official documentation notes attachments, project files, or AI-generated images cannot be imported. The update also mentions Gemini Live received a 3.1 Flash upgrade to improve conversational fluency and context retention, with continued expansion of personalized features like Personal Intelligence.

JD Cloud launched its 'Lobster Team' product line and open-sourced the instruct-tuned version of its general-purpose foundation model JoyAI-LLM Flash: 48B parameters with 3B activated. Reports indicate the model leverages FiberPO reinforcement learning and 'fiber bundle' mathematical tools to enhance computational and coding abilities. Alongside配套 OpenClaw lightweight cloud instances, appliances, and SaaS offerings, user growth exceeded 300% in one week, with token usage up 455% month-over-month. JD also introduced the digital human framework JoyAvatar and aims to accumulate 5 million hours of real-world video data for embodied intelligence within one year.