Xiaomi has released the MiMo-V2 series, featuring three self-developed models: the text base model MiMo-V2-Pro with over 1T total parameters and 42B activated, supporting up to 1 million tokens of context for complex Agent tasks; alongside a multimodal model MiMo-V2-Omni and an emotional TTS model. The Pro and Omni models have opened APIs: $1 per million tokens input and $3 output within 256K; rates double within 1M. Lei Jun announced that Xiaomi's AI investment will exceed 16 billion yuan by 2026.

OpenAI announced its plan to acquire Python developer tools company Astral, integrating the team into the Codex development line to enhance automation capabilities in dependency management, code inspection, and type checking for its cloud-based coding agent. OpenAI disclosed that Codex now has over 2 million weekly active users, a roughly threefold increase since early 2026. The deal amount was not disclosed and remains subject to regulatory and standard closing conditions. OpenAI stated it will continue maintaining Astral's open-source tools and community contributions post-acquisition.

According to internal records cited in a report, during an internal technical support session at Meta, an AI agent exposed restricted company and user data in its response, with unauthorized access lasting approximately two hours—classified as the highest severity level, Sev 1. Subsequent analysis indicated the agent could be misled into invoking privileges despite holding valid credentials, representing a 'confused deputy' risk: traditional identity verification lacks intent validation and runtime constraints after authentication. The incident highlights the need for enterprises to establish agent inventories, short-term credentials, and auditable tool invocation chains.

HackerOne has launched its 'Agentic Prompt Injection Testing' service, conducting real-environment, multi-round adversarial testing on deployed LLM applications, covering end-to-end attack paths such as indirect prompt injection, tool misuse, and data leakage. It delivers reproducible evidence and severity assessments to help enterprises verify exploitability. HackerOne reported a 540% year-over-year increase in verified prompt injection vulnerabilities, emphasizing that static filtering or runtime interception alone is insufficient. Once systems integrate internal data and executable tools, agent applications should be treated as a new attack surface requiring continuous red teaming and regression testing.

EverMind has unveiled the MSA memory architecture, claiming end-to-end support for 100 million token context lengths, combining sparse attention, document-level RoPE, KV cache compression, and memory parallelism, enhanced with Memory Interleave for multi-hop reasoning. Experimental results show less than 9% performance drop when scaling context from 16K to 100M tokens. The team also provides an open-source implementation, stating it can run on two A800 GPUs. This approach aims to shift ultra-long memory from external RAG retrieval to built-in model capability, though current disclosures are primarily from team announcements and demos.

TrendForce forecasts global foundry revenue to grow 24.8% year-on-year in 2026, reaching approximately $218.8 billion, driven by sustained procurement of AI GPUs and custom AI chips from North American CSPs and AI startups. Capacities for 5/4nm and more advanced nodes are expected to remain fully loaded. The report notes TSMC has raised prices for 5/4nm and below nodes, while Samsung has informed clients of planned price increases for similar advanced processes in 2026. For mature nodes, demand remains affected by uncertainties in consumer electronics and cost pressures, making full utilization difficult and widespread price hikes unlikely.

Universal Robots and Scale AI jointly launched UR AI Trainer at GTC, focusing on collecting high-fidelity, multimodal synchronized data (vision, force control/tactile, etc.) on real production-line hardware via human-robot demonstration, to train Vision-Language-Action models and shorten the 'lab-to-factory' deployment gap. The solution leverages UR's collaborative robots' torque control and force feedback capabilities, with Scale providing data and training platforms to create an iterative data flywheel. Both parties plan to release a large-scale industrial dataset collected using UR robots later this year, aiming to advance embodied intelligence training and deployment in industrial scenarios.

Cursor has introduced its in-house coding model, Composer 2, designed for long-chain agentic programming tasks, improving performance on benchmarks including CursorBench, Terminal-Bench 2.0, and SWE-bench Multilingual through continued pretraining and reinforcement learning. Pricing is set at $0.50 per million input tokens and $2.50 output for the standard version; a faster variant costs $1.50 input and $7.50 output and is now the default. Cursor states the model can handle complex coding tasks requiring hundreds of steps, and users can try it in Cursor and its new interface's early alpha to improve efficiency in end-to-end code modification, command execution, and debugging.

LangChain has launched LangSmith Fleet, an enterprise-grade agent management platform addressing 'management challenges at scale,' including ownership, identity, permissions, and auditing. Fleet distinguishes agent identities into shared service accounts (Claws) and user-representative Assistants (OAuth), differentiating fixed credentials from user-level authorization. It introduces permission and sharing controls to prevent uncontrolled proliferation of 'shadow agents' and adds an Agent Inbox for centralized handling of human approvals, rollbacks, and task delegation. Combined with native tracing, the platform enables tracking of every tool invocation and data access path, providing audit trails for compliance checks and incident reviews.

During discussions on the AI Omnibus proposal, the European Parliament's LIBE and IMCO committees proposed extending transition periods for certain key obligations under the AI Act, citing delays in technical standardization and the need for clearer implementation timelines and compliance planning for enterprises. Members also advocated banning 'nudifier' tools—applications using generative AI to synthesize non-consensual nude images of real individuals—to counter privacy violations and personal harm caused by deepfake abuse. These amendments still require trilogue negotiations among the European Parliament, Council, and Commission before finalization, leaving compliance timelines uncertain.