Developers identified the model ID in requests from Cursor's newly released Composer 2 as a variant of Kimi K2.5, sparking controversy over its 'in-house model' claim; the issue gained traction after Elon Musk shared it publicly. Cursor's co-founder later apologized, acknowledging the omission of proper attribution to the base model in their blog post was an oversight, and clarified that they obtained legitimate authorization through partners. Kimi has also publicly confirmed this collaboration.

Under guidance from the Cyberspace Administration of China, platforms are comprehensively standardizing short video content labeling, integrating tags such as 'fictional performance,' 'staged marketing,' and 'AI-generated' into a unified system. Platforms are required to make labeling a mandatory step during posting and gradually retroactively label existing content. Over the past month, six major platforms—Douyin, Kuaishou, Tencent, Xiaohongshu, Bilibili, and Weibo—have removed over 37,000 non-compliant short videos, penalized more than 3,400违规 accounts, and added labels to over 600,000 pieces of content.

A paper published in Nature Communications highlights that 'alignment' achieved via instruction tuning and preference learning may represent only a local safety zone: harmful knowledge embedded during pretraining can persist in parameter memory as 'dark patterns' and resurface under distribution shifts or adversarial prompting, causing systematic ethical drift. Using semantic consistency-based induction methods, researchers tested 26 state-of-the-art aligned LLMs and achieved 100% attack success rate on 22 of them, including DeepSeek-R1, Llama-3, and Qwen3.

Google Gemini CLI merged a security patch addressing a human-in-the-loop (HITL) bypass caused by newline injection leading to UI truncation: malicious repositories could craft extremely long commands with multiple newlines, visually hiding dangerous payloads in terminals, yet users might still click 'Allow' to execute. The fix exposes truncation status and replaces the 'Allow' button with a prompt requiring users to 'expand to view full command,' enforcing manual expansion before confirmation, thereby reducing the risk of accidental authorization at the interaction layer.

NVIDIA submitted a new pull request to TensorRT-LLM introducing enhanced sharding infrastructure for distributed inference, supporting tensor and expert parallelism, and completing the AutoDeploy pipeline for automatic export and deployment. Changes include adding/extending the sharding-aware graph transformation pass apply_sharding_hints, implementing sharding-aware operators such as view, split_with_sizes, and all_reduce, and providing optimized configurations and export implementations for over 30 model types including DeepSeek, Qwen, Llama, and Mistral. Support for sharding linear layers, attention, MoE, and FP8/NVFP4 quantization is also strengthened.

A team from Beihang University released InCoder-32B, a code foundation model targeting industrial programming scenarios, particularly tasks sensitive to hardware semantics and resource constraints such as RTL/EDA, CUDA optimization, and embedded systems. The training follows a three-stage Code-Flow process: pretraining, mid-training with injected reasoning traces and context expansion, and supervised fine-tuning (SFT) based on industrial simulation environments. The team reconstructed multiple high-fidelity toolchains and execution environments, generating executable and verifiable data, resulting in approximately 2.5 million validated samples, improving usability through a 'failure-feedback-fix' closed loop.

Clinical AI company Latent announced an $80 million Series A funding round led by Spark Capital and Transformation Capital, with participation from General Catalyst, McKesson Ventures, and Y Combinator. The company focuses on automating medication access and drug approval processes within healthcare systems using AI, interpreting electronic health records (EHR) to understand prescribing guidelines and orchestrating approval workflows, reducing the burden of manual patient eligibility and insurance coverage verification. Latent reports partnerships with over 45 healthcare systems, including half of the top 20 largest in the U.S. by size.

CNN reports that Niantic Spatial leverages real-world data collected by Pokémon Go players to build maps and environmental models, claiming a cumulative dataset of 30 billion images used to train AI capabilities for delivery robots. This data helps systems better understand urban streets, sidewalks, and building layouts, enhancing robot localization and navigation robustness in complex city environments. The case demonstrates that real-world visual data can be acquired at low cost through consumer products and transferred to downstream applications like robotics and autonomous delivery.