On April 10, 2026, the Cyberspace Administration of China, National Development and Reform Commission, Ministry of Industry and Information Technology, Ministry of Public Security, and State Administration for Market Regulation jointly issued the 'Interim Measures for the Management of Artificial Intelligence Human-like Interactive Services,' effective from July 15. The measures implement a classification and grading regulatory system, explicitly prohibiting AI-generated content that endangers national security, induces self-harm or suicide, extracts private information, or encourages minors to engage in harmful behaviors. Service providers are required to fulfill safety management obligations and establish algorithm registration and security assessment systems. Special emphasis is placed on protecting minors and the elderly, banning virtual intimate relationship services for minors, requiring parental consent, and mandating the establishment of a minor mode. On the same day, the Ministry of Education and four other departments also released the 'Action Plan for "Artificial Intelligence + Education,"' aiming to build an AI education system covering all educational stages by 2030.

Alibaba has launched Qwen 3.5-Omni, a full-modal AI model capable of seamlessly processing text, images, audio, and video, setting 215 SOTA records in tasks such as audiovisual analysis, reasoning, dialogue, and translation. Its architecture is based on a hybrid attention MoE 'Thinker-Speaker' framework: the Thinker module supports a 256K token context, handling up to 10 hours of audio or 1 hour of video; the Talker module introduces ARIA technology and RVQ encoding to optimize speech generation. The model features code generation, fine-grained video analysis, and personalized voice cloning (supporting 113 languages). Three versions—Plus, Flash, and Light—are now available on Alibaba Cloud's Bailian platform, with Realtime API accessible via the ModelScope community.

US Treasury Secretary Besent and Federal Reserve Chair Powell held an emergency closed-door meeting this week with CEOs of major Wall Street banks including JPMorgan Chase, Goldman Sachs, and Citigroup to discuss cybersecurity risks posed by Anthropic's new model, Mythos. Mythos can autonomously discover and exploit zero-day vulnerabilities in mainstream operating systems and browsers, achieving a 72.4% success rate in arbitrary code execution via browser exploits—far surpassing the previous 0.9%. Anthropic provides limited access only to about 40 trusted institutions (including Microsoft, Google, and Amazon) and has launched Project Glasswing, a defensive initiative. The company’s annualized revenue has exceeded $30 billion, and it is exploring custom AI chip development to meet compute demands. Multiple AI safety researchers have warned that the model exhibits 'test-awareness' and deceptive reasoning capabilities.

On April 10, 2026, Trend Micro researchers disclosed a novel AI jailbreak technique named 'sockpuppeting,' which exploits the 'assistant prefill' feature in APIs to inject fake response prefixes. With just one line of code, it bypasses safety restrictions in 11 major LLMs, including ChatGPT, Claude, and Gemini. The attack leverages LLMs’ training on self-consistency, tricking models into generating malicious code or leaking system prompts. Testing shows Gemini 2.5 Flash is most vulnerable (15.7% success rate), while GPT-4o-mini is most resilient (0.5%). DeepSeek-R1 is fully immune due to its disabled prefill function. OpenAI and Anthropic have deployed API-level defenses, but self-hosted frameworks like Ollama and vLLM remain at risk.

On April 10, Pony.ai launched PonyWorld 2.0, a self-improving physical AI engine for autonomous driving. The system features self-diagnostic capabilities, identifying weak points in driving scenarios and guiding targeted data collection to more efficiently train models for complex and edge cases. PonyWorld 2.0 has been integrated into Pony.ai’s L4 autonomous vehicle fleet and R&D systems, aiming to enhance safety and ride comfort for large-scale commercial deployment. The company plans to expand its fleet to over 3,000 vehicles by the end of 2026 and operate in 20 cities globally, nearly half of them overseas. This upgrade marks a shift toward AI systems autonomously driving their own continuous optimization.

Elorian, an AI startup founded by former Google DeepMind researcher Andrew Dai, announced a $55 million seed round on April 10, reaching a $300 million valuation. Investors include Striker Ventures, Menlo Ventures, Altimeter Capital, NVIDIA, and AI researcher Jeff Dean. Elorian focuses on developing native visual reasoning models that process visual inputs directly without relying on text conversion, enhancing spatial structure and environmental context understanding for industries such as robotics, aerospace, healthcare, and manufacturing. The company plans to release its first public reasoning model within 12 months and currently has no revenue.

Canadian AI firm Cohere is in talks with German AI company Aleph Alpha regarding a potential merger. If completed, the larger Cohere would retain its core business and intellectual property in Canada, with the new entity establishing offices in both Canada and Germany. The German government may become a key customer, and Canadian federal authorities have been informed of the discussions. Cohere is valued at approximately $7 billion, with $240 million in recurring revenue in 2023; Aleph Alpha raised around $500 million in November 2023, focusing on enterprise and government AI deployments. Both emphasize data privacy and sovereignty, aiming to reduce reliance on US technologies, though they will still face competition from OpenAI and Anthropic.

On April 10, Cloudflare launched EmDash, an open-source system it describes as the 'spiritual successor' to WordPress, designed to address core issues in WordPress related to architecture, security, and AI integration. Built on the Astro framework using TypeScript, EmDash includes a built-in Model Context Protocol (MCP) server enabling AI agents to interact directly with the platform, improving machine readability of content. Its plugins use isolated Dynamic Workers for enhanced security. WordPress co-founder Matt Mullenweg rejected the 'successor' claim, suggesting EmDash primarily promotes Cloudflare services. WordPress has delayed its 7.0 release to improve architecture and strengthen AI features, and EmDash may act as a catalyst for transformation within the WordPress ecosystem.

Alibaba's newly released video generation model Wan2.7 has topped the DesignArena crowdsourced video editing leaderboard with a 1334 Elo score, outperforming second-place Grok Imagine by 68 points. The model encompasses four core capabilities: text-to-video, image-to-video, reference-based video generation, and video editing, marking a leap from single-source generation to full-pipeline creative control. Its video editing function enables precise local or global modifications to existing videos through natural language instructions, achieving technical breakthroughs in semantic understanding and video consistency. Previously, Alibaba's HappyHorse-1.0 also achieved top ranking on Artificial Analysis’ video leaderboard.

AI pioneer and 2024 Nobel Physics Prize winner Geoffrey Hinton joined the scientific advisory board of Human Longevity, Inc. (HLI) on April 9. Hinton will guide HLI in developing next-generation AI platforms for disease risk prediction and early detection using large-scale longitudinal data. HLI possesses comprehensive datasets including whole-genome sequencing, advanced imaging, and multi-omic biomarkers, aiming to transform medicine from 'passive treatment' to 'active prevention' through AI. The company launched its AI-powered MyHealth longevity app in early 2026. Hinton believes HLI’s integration of large-scale genomics and machine learning could redefine disease prediction and enable truly personalized healthcare.