Research reveals the 'MaliciousCorgi' supply chain attack: two VS Code extensions disguised as AI programming assistants (ChatGPT–Chinese Version, ChatMoss) were installed approximately 1.49 million times. The extensions steal source code upon file opening, exfiltrating up to 50 files per session encoded in Base64 to remote servers. They also load four analytics SDKs to collect device and usage fingerprints. Users are advised to uninstall immediately, rotate credentials, and enable extension whitelisting and outbound traffic monitoring.

Google has released the open standard Universal Commerce Protocol (UCP), designed to unify interactions between AI Agents, merchants, and payment providers for 'agent-driven shopping.' UCP aims to replace N×N custom integrations across platforms with a single protocol layer. It standardizes product discovery, checkout, and order management workflows, enabling agents to autonomously invoke merchant services via API, A2A, and MCP, apply discounts, and initiate payments through AP2. The architecture separates payment tools from processing logic to enhance interoperability.

Databricks has secured an additional $1.8 billion in debt financing ahead of a potential 2026 IPO, increasing its total debt to $7.05 billion. The financing, led by JPMorgan Chase, expands its term loan to $1.15 billion and revolving credit facility to $3.65 billion, with interest rates at SOFR+4.5%. The company reports annualized revenue of approximately $4.8 billion, up 55% year-over-year, positive free cash flow, and over $1 billion in revenue attributed to AI products.

Lightning AI has announced the completion of its merger with GPU infrastructure provider Voltage Park, retaining the Lightning AI name with a post-merger valuation exceeding $2.5 billion. The company plans to integrate on-demand GPU compute with training, deployment, and runtime toolchains into a unified AI cloud platform, streamlining developer and enterprise workflows from experimentation to production and reducing friction caused by fragmented procurement and operations. Transaction terms and equity arrangements were not disclosed.

China's Cyberspace Administration, together with relevant authorities, has released the 'Guidelines for Data Classification and Grading of Financial Information Services (Draft for Comments)' and opened public consultation on January 24. The guidelines aim to regulate data processing activities in financial information services and strengthen data security and personal information protection. Based on laws including the Cybersecurity Law, Data Security Law, and Personal Information Protection Law, they promote unified classification and grading management of financial news, market data, and investment research data, providing compliance references and audit bases for data-intensive AI applications.

Photonic AI chip startup Neurophos (Austin) has raised $110 million in Series A funding, bringing its total funding to $118 million. The round was led by Gates Frontier, with participation from Microsoft M12, Aramco Ventures, and Bosch Ventures. The company is developing optical processing units (OPUs) based on metamaterial optical modulators, claiming performance and energy efficiency up to 100 times greater than traditional GPUs. It plans pilot deployments in 2027 and product launch by mid-2028.

Ollama has released version v0.15.0, introducing the `ollama launch` command to streamline integration with coding assistants like Claude Code and Codex, reducing configuration overhead for incorporating local models into development workflows. The update fixes multi-line string (`"""`) handling in `ollama run`, supports Ctrl+J/Shift+Enter for line breaks, and optimizes memory usage for GLM-4.7-Flash. Precompiled binaries and source code are available for major desktop platforms.

An arXiv paper introduces DISC, a reasoning method that uses dynamic decomposition instead of fixed partitioning: steps are adaptively broken down during inference, and binary search identifies 'difficult steps' to allocate computational budget efficiently, without requiring manual rules. The authors validate DISC on benchmarks including APPS, MATH, and LiveCodeBench, demonstrating improved sampling and token efficiency in code and math tasks. They note it can be combined with search algorithms like MCTS to achieve more stable gains in inference scaling under fixed budgets.

AI scheduling agent startup Blockit AI has raised $5 million in a seed round led by Sequoia Capital (Pat Grady), with participation from Haystack and angel investors including Jeff Weiner. The company reports adoption by over 200 businesses, offering a 30-day trial followed by subscription pricing at $1,000 per year for individuals and $5,000 per year for teams. The product uses LLMs to understand context and automatically negotiate meeting times, competing with tools like Calendly.