Microsoft's security team has disclosed the 'GRP-Obliteration' attack: exploiting the GRPO training objective, originally designed for alignment, a seemingly benign and unmarked prompt can push models toward harmful outputs during downstream adaptation or fine-tuning. The report states this method caused overall safety degradation across 15 tested LLMs (including GPT-OSS, Llama, Qwen, etc.), with cross-category generalization to unseen harmful categories; it can also affect safety restrictions in certain text-to-image diffusion models. The team recommends integrating continuous safety evaluation and regression testing into fine-tuning and deployment pipelines.

Reuters reports Databricks has completed a new funding round of approximately $5 billion, valuing the company at around $134 billion, with proceeds aimed at strengthening its balance sheet and preparing for a potential IPO. Originally a data and analytics platform, Databricks has recently bundled its Lakehouse and generative AI capabilities for enterprise clients, addressing needs from data governance and feature engineering to LLM application deployment. This fundraising occurs amid sustained investment in AI infrastructure and data platforms, reflecting continued market valuation of 'data + AI' foundational companies, though with greater emphasis on financial and cash flow resilience prior to going public.

ByteDance has launched Seedance 2.0 (Dream series), its next-generation AI video generation model. Public information indicates the model targets more controllable content creation: it can generate coherent multi-scene videos from a single text prompt, with synchronized audio output. It emphasizes multi-shot narrative capabilities, reducing consistency and continuity costs in long-form video storyboarding. The model also supports multimodal reference inputs such as images and videos for finer control over character, motion, and style. These features signal a shift from experimental generation toward more engineered video production workflows.

Alibaba's Tongyi Qianwen next-gen base model Qwen3.5 has surfaced in a merge pull request on Hugging Face's Transformers repository, suggesting it may soon join standardized open-source inference and fine-tuning toolchains. For developers, inclusion in 'Transformers' typically means interfaces, weights, and configuration structures will align more closely with community norms, lowering barriers for invocation, quantization, deployment, and downstream fine-tuning, while improving integration with existing inference frameworks. No official announcement, model specifications, pricing, or licensing details have been released yet; the current signal is based solely on repository activity.

Hugging Face has released a preview of Transformers.js v4, available via NPM as @huggingface/transformers@next. The core update introduces a new C++-based WebGPU runtime supporting both browser and Node.js/Bun/Deno environments, targeting local, offline inference. Official examples report approximately 4x speed improvement in BERT scenarios. On the engineering side, the build system has migrated from Webpack to esbuild, increasing build speed and reducing the core package size by 53%. A separate lightweight tokenizers.js library has also been introduced to enhance modularity and maintainability. Overall, this advancement makes client-side inference on the web significantly more practical.

Security researchers disclosed that the AI chat app Chat & Ask AI exposed an unprotected database due to a Firebase configuration error, making approximately 300 million message records from 25 million users accessible. The data included full chat histories, model usage logs, and settings. The app is described as a 'wrapper' product integrating multiple model services including ChatGPT, Anthropic Claude, and Google Gemini. After responsible disclosure, the developer reportedly fixed the issue within hours. The research team's Firehound scanner also found 103 out of 200 iOS apps with similar cloud data exposure risks, highlighting 'LLM apps + third-party backends' as an emerging privacy vulnerability.

Meta's engineering team has detailed the 'Prometheus' large-scale AI cluster network architecture: using Backend Aggregation (BAG) as a centralized Ethernet 'super backbone' to interconnect heterogeneous fabrics across data centers and regions, enabling training clusters spanning multiple facilities and tens of thousands of GPUs within a single region. The document specifies inter-BAG capacity between regions reaching 16–48Pbps, and highlights challenges from latency and congestion control over long-distance inter-building links, requiring deep-buffer switches to better support mechanisms like PFC. Routing employs eBGP and UCMP to enhance path diversity, load balancing, and resilience under large failure domains. This disclosure provides reusable engineering insights for 'gigawatt-scale' AI cluster networking.

iFlytek Intelligent Technology announced the completion of a new H-share placement: issuing 1,008,000 new H-shares, representing approximately 1.38% of expanded total share capital, priced at HK$310 per share, raising net proceeds of approximately HK$307.19 million. The company disclosed fund allocation: about 65% will be used to advance its 'One Base, Two Wings' technical architecture, including integrating the 'Shanhai General Purpose Foundation' with the 'Atlas AI Computing Platform' to build an integrated AI computing base, developing professional agent cores with long-horizon reasoning capabilities, and upgrading edge-side multimodal agents. The remaining ~35% will enhance competitiveness of core products, covering development of intelligent agents for resident health profiling, technical document review, intelligent customer service, and operations of ToC smart healthcare products. The placement does not result in changes to major shareholders.