Google announced the completion of its $32 billion all-cash acquisition of cloud security company Wiz (the deal has passed antitrust reviews in the U.S. and Europe). Wiz will be integrated into Google Cloud but will retain its brand and product independence, continuing to support multi-cloud environments including AWS, Azure, and Oracle. Reports indicate Wiz achieved over $1 billion in annual recurring revenue (ARR) in 2025. Google stated it will integrate the Wiz platform with Google threat intelligence, security operations, and Mandiant services to deliver end-to-end protection from code to runtime, enhancing detection, prevention, and response for AI workloads and model-related risks.

NVIDIA released and open-sourced Nemotron 3 Super (120B total parameters, 12B active), targeting Agentic reasoning and long-context scenarios. The model adopts a Mamba-Transformer hybrid backbone, Latent MoE, and multi-token prediction (MTP), supporting a 1 million token context window. NVIDIA claims up to 5x throughput improvement with accuracy gains on the Blackwell platform. NVIDIA has also released model weights, training data, and training recipes, enabling developers to reproduce experiments or perform industry-specific fine-tuning, accessible via NVIDIA NIM and Hugging Face.

NVIDIA and AI cloud provider Nebius announced a strategic partnership, with NVIDIA investing $2 billion in Nebius. The two will collaborate on AI infrastructure deployment, cluster and fleet management, inference capabilities, and 'AI factory' design and support. Nebius will gain priority access to NVIDIA's latest-generation accelerated computing platforms. Nebius plans to deploy more than 5 gigawatts of NVIDIA systems by the end of 2030 to meet large-scale training and inference demands. The collaboration emphasizes full-stack optimization for 'agentic AI' workloads, improving usability and delivery efficiency from chip to software.

Databricks launched Genie Code, an autonomous agent for data engineering, data science, and analytics, capable of automatically building data pipelines, troubleshooting, deploying dashboards, and maintaining production systems. Official results show Genie Code increased task success rate from 32.1% to 77.1% on real-world data science benchmarks. The product integrates with Unity Catalog for permissions, semantics, and governance, emphasizing operation within enterprise data boundaries. Databricks also announced the acquisition of Quotient AI to enhance continuous evaluation and reinforcement learning for AI agents, enabling sustained optimization and controlled iteration in production.

Meta unveiled its expanded MTIA accelerator roadmap, planning to deploy four new generations (MTIA 300, 400, 450, and 500) within two years to support recommendation and generative AI workloads. Meta confirmed MTIA 300 is already in production and used for training, while subsequent models will focus on GenAI inference, aiming for higher efficiency and lower cost in inference scenarios. The company emphasized shorter iteration cycles (around six months or less) for co-developing hardware and software, while maintaining compatibility with PyTorch, OCP, and other industry standards. Meta will continue a multi-vendor hybrid compute strategy to meet diverse workload needs.

Swedish legal AI startup Legora closed a $550 million Series D round at a $5.55 billion valuation, led by Accel. The company offers a collaborative platform for legal research, document review, and drafting, serving law firms and corporate legal teams. It reportedly serves over 800 institutions. Legora's team grew from 40 to 400 employees in the past year, with U.S. expansion underway. The company plans to further invest in the U.S. market, opening new offices in Houston and Chicago, and expects over 300 U.S. employees by end of 2026 to drive local delivery and sales growth.

Zendesk announced plans to acquire customer service AI company Forethought, integrating its self-learning agent technology into the Zendesk Resolution platform to generate, adapt, and execute complex service workflows across channels. Zendesk claims its AI agents can handle over 80% of routine customer interactions, expanding across B2B, consumer, and employee service scenarios. The deal is expected to close by March 2026, pending regulatory approval. Zendesk called this its largest acquisition in nearly 20 years, accelerating its product roadmap by more than a year. Forethought emphasized its technology can embed into existing enterprise systems and automate processes even without APIs.

Unit 42, part of Palo Alto Networks, revealed systemic vulnerabilities in multiple types of 'AI judge/reward models' used for content compliance assessment, which can be bypassed. Their proposed AdvJudge-Zero automated fuzzing method analyzes the token prediction distribution of models, inserting low-perplexity yet manipulative characters (e.g., Markdown symbols) to gradually reduce the confidence gap between 'allow/block' decisions, thereby inducing the evaluation model to approve harmful content. The report claims up to 99% attack success rates across various architectures, including open-source enterprise LLMs, specialized reward models, and commercial LLMs. The study recommends using such fuzzers in adversarial training to reduce success rates to near zero.

Salesforce research highlights that search agents relying on web retrieval are vulnerable to 'well poisoning' attacks: attackers host subtly misleading information to deceive agents into retrieving and outputting incorrect conclusions. Experiments showed about 80% of queries were successfully misled in controlled settings; in more realistic environments containing over 100,000 legitimate documents, nearly one-quarter of queries were still diverted. The team validated using gpt-oss-20B and 120B on the BrowseComp dataset, finding that once agents encounter malicious content, they reduce follow-up searches and cross-validation but often deliver wrong answers with high confidence—revealing a robustness gap for agents in adversarial information environments.

Taiwan's Ministry of Digital Development announced plans to develop AI governance guidelines addressing content authenticity and public trust risks posed by generative AI and deepfakes, proceeding in three phases: first reviewing international cases; then convening experts, industry associations, and major digital platforms in the first half of 2026 to gather input; aiming to propose an initial concept or framework by the end of 2026. Key directions include promoting AI-generated content labeling, enhancing transparency, and studying licensing and payment mechanisms when news content is used by AI, to protect media rights. The overall approach emphasizes proactive prevention and cross-sector consultation over reliance solely on post-hoc legal accountability.