Automated AI News Brief: Hy4, Copilot, and RAG Access Control
August 30 AI news brief: Tencent Hy4 Preview, OpenAI and Cursor, GitHub Copilot updates, a Thailand AI startup accelerator, and a RAG access-control checker.
Introduction
Horizon collected the source material for this post, while Codex selected and rewrote it. Horizon is used only for data collection.
1. Tencent Releases the Open-Weight Hy4 Preview
Tencent's Hy4 Preview is an open-weight, text-input-only LLM. Simon Willison's roundup of the official release describes 770 billion total parameters, 49 billion active parameters, a one-million-token context window, and model files available on Hugging Face. For teams considering self-hosting or evaluating a large open model, hardware, inference infrastructure, and task-specific validation remain the practical constraints, not merely access to the weights.
Sources: Tencent: Hy4 Preview, Simon Willison: Introducing Hy4 Preview, Hugging Face: Hy4 Preview
2. OpenAI Will Wind Down Its Contract to Provide Models to Cursor
OpenAI says it will wind down its contract to provide OpenAI models to Cursor following Cursor's acquisition by SpaceX. The Horizon-captured announcement does not provide further implementation details. For development teams that rely on a single model supplier inside an IDE agent, it is a useful prompt to inventory model dependencies, alternatives, and billing arrangements before a business relationship changes.
Source: OpenAI: Our decision on Cursor following its acquisition by SpaceX
3. GitHub Copilot Expands Controls Across Slack, Teams, CLI, and IDEs
GitHub's weekly Copilot updates focus on giving teams more control over how Copilot runs, from team sessions in Slack and Teams to customization in the app, CLI, and IDEs. The agent-adoption question is increasingly shifting from whether a tool works to who may use it under which settings and how those settings stay consistent. Teams already using Copilot broadly should make shared configuration and collaboration rules explicit and reviewable.
Source: GitHub Changelog: GitHub Copilot weekly releases — August 24
4. OpenAI and Thailand's MHESI Launch an AI Startup Accelerator
OpenAI and Thailand's Ministry of Higher Education, Science, Research and Innovation (MHESI) have launched an eight-week accelerator for 10 health, wellness, and education startups, aimed at turning AI prototypes into more trusted products. This is not a product-feature launch, but it shows how model companies are building application ecosystems through local partners and verticals with clear trust requirements.
Source: OpenAI: Supporting Thailand's next generation of AI startups
5. Community Project Releases a RAG Access-Control Checker
An open-source community tool says it checks whether a RAG application retrieves documents that a user should not be able to access. It supports offline test cases and live HTTP API tests using bearer-token or API-key authentication. This is a community project's own description, not a general security guarantee, but it highlights an often-missed RAG concern: retrieval permissions need practical tests rather than relying only on prompts or interface-level access design.
Sources: Reddit: Open-source access-control checker for retrieval-based AI applications, GitHub: rag-access-check
Takeaway
Today's news moves both ends of the AI toolchain: larger open models and broader agent collaboration surfaces continue to emerge, while vendor relationships, team governance, and retrieval permissions demand more deliberate design. Capability is expanding quickly; durable workflows still depend on alternatives, shared settings, and verifiable access control.