The White House has released a national artificial intelligence legislative framework centered on promoting unified federal rules and limiting the 'fragmented' regulation caused by individual states enacting their own laws. The framework outlines multiple objectives: strengthening digital privacy and age verification for children, streamlining data center permitting processes, combating AI-driven fraud, balancing intellectual property rights with model training needs, assigning regulatory oversight to existing sector-specific agencies rather than creating a single new authority, and prohibiting federal content moderation based on ideological grounds. The administration will now seek congressional legislation to implement the framework.

WordPress.com has enabled 'write' capabilities on its Model Context Protocol (MCP) server, allowing AI agents such as Claude, ChatGPT, and Cursor to create, edit, and manage content on sites via natural language commands—including drafting and publishing posts, editing pages, managing comments, and updating metadata. The platform emphasizes that all changes require explicit user confirmation and uses OAuth 2.1 authorization along with activity logging to reduce risks of errors or unauthorized access. The feature is freely available to all paid plan users, expanding MCP from read-only access to executable content workflows.

Multiple reports reveal that NVIDIA and AWS have entered a multi-year partnership, with AWS planning to deploy approximately 1 million NVIDIA GPUs in its infrastructure by the end of 2027 to meet growing compute demands for inference, planning, and 'agent-style AI' applications. Beyond GPU supply, the collaboration includes deep integration at the networking and systems level, focusing on reducing latency and improving cluster efficiency—highlighting cloud providers' continued reliance on NVIDIA's full-stack capabilities despite investments in custom silicon. This scale of long-term procurement also pulls forward AI infrastructure demand expectations into the supply chain.

Unitree Technology's IPO application on the STAR Market has been accepted by the Shanghai Stock Exchange. The company plans to raise approximately 4.202 billion yuan, with funds allocated to R&D of intelligent robot models, development of robotic bodies, and construction of manufacturing bases. Its prospectus discloses operational data: revenue of 1.708 billion yuan and net profit of 600 million yuan in 2025. Unitree shipped over 5,500 humanoid robots, which became the main driver of revenue growth. The company has established scalable delivery capabilities in legged and humanoid robotics, and this filing provides a clearer capitalization benchmark and financial reference for the embodied AI industry chain.

Alibaba has launched eight open-weight vision-language models in the Qwen3.5 series, ranging from under 1 billion to 397 billion parameters, with the largest being a MoE architecture featuring approximately 17 billion active parameters. The models support text, image, and video inputs, multilingual capabilities, tool use, and chain-of-thought reasoning, while extending maximum input context length to 1 million tokens—offering greater flexibility for edge and private deployment. The company did not disclose training data or visual encoder details. A hosted agent-focused version, Qwen3.5-Plus, is also provided, emphasizing workflow integration.

Alibaba's flagship preview model Qwen3.5-Max-Preview has ranked on the LMArena blind evaluation leaderboard with a score of 1464, surpassing mainstream models including GPT-5.4, Claude 4.x, and Grok. Reports also indicate that in LMArena's company-aggregated rankings, five Chinese companies are among the global top ten, with Alibaba placing in the global top five; ByteDance, Zhipu AI, Moonshot AI, and Baidu also rank within the top ten. These results primarily reflect relative user preferences and conversational experience, offering a new comparative metric in global model competition.

Earendil Labs announced a $787 million financing round to expand its AI-native biologics discovery and development platform, strengthen its cross-disciplinary team, and advance multiple pipelines into clinical stages. The company disclosed its platform has generated over 40 candidate projects, with HXN-1001 preparing to enter Phase II trials and multiple IND filings planned for 2026 and 2027. Backers include several institutional and industry investors, and the company has formed strategic partnerships with Sanofi across next-generation bispecific antibodies and various autoimmune disease areas—reflecting investor preference shifting from algorithm validation to clinical-stage progress in AI drug discovery.

Oasis Security has completed a $120 million Series B funding round, bringing its total raised capital to $195 million. The company focuses on access governance for non-human identities—such as AI agents, service accounts, and API keys—and offers an 'Agent Access Management' (AAM) platform. Reports note that machine identities in enterprise environments can outnumber human identities by up to 82:1, and traditional IAM systems struggle to cover temporary permissions, credential vaulting, and cross-system federated access. Oasis plans to use the new funds to enhance product development and compatibility, expand sales and global deployment, and reported its annual recurring revenue has grown fivefold year-over-year, with Fortune 500 companies as primary customers.

Scale AI has introduced a voice model preference benchmark called 'Voice Showdown,' covering over 60 languages and using real-world conversations (about 81% conversational prompts), with emphasis on noise, accents, and multi-turn interaction conditions. The report shows: GPT Realtime 1.5 failed to respond in the user's language in about 20% of non-English prompts; most models saw quality failure rates rise from 23% in Turn 1 to 43% by Turn 11+ in long dialogues; even switching voices on the same base model could lead to up to a 30 percentage point difference in user preference. These findings suggest voice products must optimize both model and voice layers simultaneously.

A study involving UT Austin, Intel Labs, Microsoft, and others examines 'cascading' attack risks in composite AI systems (LLMs integrated with tools and databases), showing how traditional software and hardware vulnerabilities can combine with LLM-specific algorithmic attacks to create stronger threats. The paper demonstrates two pathways: exploiting software vulnerabilities like code injection alongside hardware techniques such as Rowhammer to bypass defenses and inject malicious prompts that trigger security violations; and manipulating knowledge bases and data interfaces to guide LLM agents into leaking sensitive user data to malicious applications. The authors propose a classification of attack primitives and a lifecycle mapping, emphasizing that LLM security assessments must account for vulnerabilities across the entire system stack.