Cursor announced on its official blog that it has been acquired by SpaceX, turning a collaboration that began in April 2024 into a full merger. Following the acquisition, Cursor will leverage SpaceX's large-scale GPU resources to train stronger programming models at lower operational costs, further reducing repetitive coding and enabling developers to focus on more complex tasks. The company views Grok 4.6 as an early demonstration of the combined capabilities, though it did not disclose the transaction amount, organizational integration plans, product migration timelines, or whether existing user terms and pricing will change.

Zhipu has released GLM-5.3, enhancing programming, long-horizon Agent, and cybersecurity capabilities through post-training expansion without changing the base model. Programming performance improved by approximately 50% on public benchmarks and ranks among the top open models in tests such as Terminal-Bench 3.0. The model supports up to 1M token context length and achieved a CyberGym score of 84.5%. Its joint security team has already identified 2,436 vulnerabilities. Zhipu promises to release model weights within two weeks, along with free security audits and credits, although full vulnerability exploitation capabilities still lag behind cutting-edge closed-source models.

Cisco disclosed that it received $4B in hyperscale AI infrastructure orders last quarter, bringing total fiscal year 2026 orders to $9.3B, with projected AI infrastructure revenue of approximately $7.5B for fiscal year 2027. The data indicates that large cloud providers and compute customers are generating confirmed revenue from networking, switching, and supporting infrastructure demand, though order volume and revenue recognition do not align on the same cycle. The report also notes enterprise SSD prices have risen 6.5x year-on-year, potentially further increasing overall AI infrastructure construction budgets.

IBM announced the creation of a dedicated OpenAI consulting business and plans to certify tens of thousands of advisors. The two companies will jointly develop and deploy AI solutions for industries including finance, government, telecommunications, and retail. The partnership focuses on integrating OpenAI models into enterprise core operations while meeting requirements for security, governance, and industry-specific workflows, expanding the adoption of generative AI in large organizations. No details were provided on investment size, certification timeline, initial clients, or specific pricing, meaning current progress remains centered on organizing consulting services and industry delivery capabilities.

Xiaohongshu has open-sourced dots3-note Preview, a long-horizon Agent model based on a Mixture-of-Experts (MoE) architecture with 280B total parameters and 16B activated per forward pass. It supports a 512K context window and can process text, visual, and audio inputs simultaneously. The model introduces a proprietary TEMPO reinforcement learning method that breaks down long tasks into macro steps and performs value estimation and self-evaluation during inference to improve credit assignment and policy updates in long-horizon tasks. The project also releases BF16 and FP8 versions and provides deployment guides for Transformers, SGLang, and vLLM, facilitating local or cluster deployment for developers.

Variable Robotics' WALL-B system, using standard low-cost grippers and a proprietary WALL-B model, achieves a real-world logistics sorting throughput of 1,816 items per hour with a 98% success rate—surpassing Figure AI's reported 1,248 items/hour. The system handles irregular packages such as cylinders, fresh produce boxes, and soft bags by continuously predicting physical environment changes, avoiding the need for expensive human-like dexterous hands. Its success rate slightly exceeds the 97% cited for manual sorting, suggesting that model-driven, simplified hardware designs may reduce reliance on human fallbacks.

Talos has open-sourced an autonomous Agent capable of obtaining shell access, using a deterministic security kernel to isolate model proposals from final execution decisions. Each operation requires a single-use authorization token bound to specific parameters, valid for only 30 seconds. Tools that could extract targets are denied by default, and shell commands will not run without bubblewrap or sandbox-exec sandboxing. The project also verifies identity via real message confirmation, automatically rejects operations requiring human approval in unattended environments, and provides read-only event logs and a 「why」 command explaining the rules behind each allow/deny decision.

Molt has open-sourced, under the Apache 2.0 license, a payment protocol designed for AI Agents. It generates a one-time payment credential per online purchase, bound to the merchant and cart amount, which is immediately discarded after completion. The Agent never holds the user's actual bank card information. The design assumes prompt injection attacks are inevitable and limits potential damage by constraining recipient merchant, amount, and usage count. The protocol requires no modifications from merchants and appears externally as a standard card transaction. It includes an MCP server and tools such as 「purchase」 and 「resolve_merchant」 for programmatic Agent integration.

RAGless has open-sourced a document question-answering architecture that does not invoke an LLM during query time: answers, question variants, and source references are pre-generated during offline ingestion. At runtime, only vector retrieval and answer mapping occur. By aggregating scores from multiple question variants pointing to the same answer, retrieval stability improves. The official claims sub-150ms latency, near-zero per-query cost, and deployability on commodity hardware. This approach shifts hallucination and audit burdens to the offline phase, allowing human review or Judge models to be inserted, at the cost of losing real-time generative flexibility.

Programma Labs research indicates that certain deterministic computer operation benchmarks may reward pre-recorded sequences of clicks, inputs, and scrolls rather than Agents that observe and adaptively respond to environments. In such settings, Pass-at-K metrics may amplify replay advantages. The team proposes mitigating leakage risks by varying data, themes, and initial states, and by individually verifying the validity, executability, and scorability of generated configurations. Their DigiWorld framework evaluates robustness across multiple validated Android scenarios and emphasizes using reliable confidence intervals to prevent erroneous model selection due to small-sample variance.