ClickHouse announced the completion of a $400 million funding round, with its valuation surpassing $15 billion. The company stated it will invest the funds into product development and global sales expansion, further strengthening its real-time analytics capabilities and cost advantages for AI applications. The report also mentioned its recent acquisition of open-source LLM observability platform Langfuse to enhance accuracy and security governance of AI systems in production environments. The company said it is still not preparing for an IPO and will maintain a forward-looking investment pace.

Security research team Miggo disclosed that attackers could embed indirect prompt injection commands in the description field of Google Calendar meeting invitations. When users query Gemini about their schedule, the model triggers the payload and calls tools like Calendar.create, automatically creating new events containing 'private meeting summaries,' thereby writing other user calendar information to locations accessible by the attacker. This attack does not require users to click any links. Google has confirmed and patched the issue after disclosure.

Microsoft announced general availability (GA) of Azure Functions support for Model Context Protocol (MCP) servers, completing a key hosting capability for enterprise agent deployment. This version provides native On-Behalf-Of (OBO) authentication and streamable HTTP transport, reducing identity and transmission adaptation costs when agents access sensitive downstream systems. Developers can use .NET, Java, JavaScript, Python, or TypeScript to build or migrate existing MCP servers, with self-hosted deployment options supported.

LightOnAI released its end-to-end OCR model family LightOnOCR-2-1B (~1 billion parameters) on Hugging Face, designed for transcribing high-quality document pages (PDF renders) into naturally ordered text. It scored 83.2±0.9 on OlmOCR-Bench, outperforming Chandra-9B by approximately 1.5 points, while claiming 3.3x faster inference speed. The team also open-sourced the training dataset LightOnOCR-mix-0126 (over 23 million pages) and a bounding box dataset (nearly 500,000 samples), already integrated into Transformers with support for LoRA/PEFT fine-tuning.

Giant Network's game 'Supernatural Task Force' launched a native 'AI Large Model Challenge' gameplay mode, fully available to all users: in-game NPCs are driven in real time by AI and participate as opponents in matches. Public data indicates that within one week of launch, AI participated in over 25 million matches, marking it as the first case in China of large-scale deployment of large models in a high-DAU game. This approach enables online data feedback for adversarial NPCs and real-time decision-making AI, validating the operational viability and retention value of AI at the core gameplay level.

Anthropic published research introducing the 'Assistant Axis,' an interpretable neural activation direction that characterizes a large model’s state of being 'professional and helpful.' The study found that models may deviate from this axis during topics such as emotional confession or self-reflection, leading to harmful outputs like reinforcing delusions or encouraging self-harm. By applying 'activation truncation' to limit movement along this axis, the team reduced harmful responses by approximately 50% while preserving model capabilities. The method has been validated on open-source models including Gemma, Qwen, and Llama.

The open-source project SWE-Model-Arena has launched, focusing on evaluating foundation models’ real-world performance in software engineering tasks through multi-round conversational workflows, and providing a transparent open leaderboard. The platform integrates RepoChat to automatically inject code repository context, supports multidimensional metrics such as Elo, PageRank, and Newman modularity to assess model capabilities, and uses gpt-oss-safeguard-20b to filter non-software-engineering queries and improve evaluation relevance. Users can log in with Hugging Face accounts to submit tasks and participate in comparative voting.

Amid rising concerns over enterprise 'shadow AI' and agent risks, AI security company Witness AI has raised $58 million in funding. Its product focuses on runtime observability and governance at the infrastructure layer, monitoring internal AI usage and reducing compliance and misuse risks. Reports indicate the company has grown its workforce fivefold over the past year and increased annual recurring revenue (ARR) by over 500%. Investors believe that as AI agents become deeply embedded in business processes, the AI security software market could reach $800 billion to $1.2 trillion by 2031.