NVIDIA 오픈 시큐어 AI 얼라이언스·Nanbeige4.2·Music‑JEPA 등 최신 AI 동향

AI security advances when the industry builds in the open, together. We're introducing the Open Secure AI Alliance with industry leaders to develop new techniques and tools to safeguard software and agents. By sharing models, tooling and research in the open, we can broaden the community of defenders. Learn more about the founding members’ contributions: https:// nvda.ws/4pD8Fc5

NVIDIA AI

@NVIDIAAI

We're contributing open models, weights, data and research to the Open Secure AI Alliance. This includes NVIDIA Labs Object-Oriented Agents (NOOA): open agent harness research in the form of an object-oriented framework that reaches state-of-the-art results across software engineering, cybersecurity and reasoning benchmarks. Developers can use the findings to optimize their own harnesses and then contribute back.

"We present Nanbeige4.2-3B, a compact general agentic model with 3B non-embedding parameters" "Nanbeige4.2-3B is pretrained from scratch on 28T tokens with a Looped Transformer that reuses the layer stack to increase capacity without adding parameters" "Nanbeige4.2-3B outperforms larger models, including Qwen3.5-9B and Gemma4-12B"

model: https:// huggingface.co/Nanbeige/Nanbe ige4.2-3B … abs: https:// arxiv.org/abs/2607.22083

Music-JEPA: Learning a World Model of Sound from Action "we propose to learn a world model of piano sound using JEPA by framing music as an action-conditioned system: the audio is treated as the state, and the pianoroll as the instrument action." "Experiments show that the learned model captures the relationships between musical actions and their resulting sound. The resulting representations support downstream tasks, including beat tracking, composer identification, and key estimation, and enable piano transcription via planning, by searching for actions that best explain a target sound."

demo: https:// zzwaang.github.io/music-jepa-dem o/ … abs: https:// arxiv.org/abs/2607.22000

The top Claude Code CLI integrations for AI engineers: (bookmark this) 1. GitHub Helps the agent read and write issues, PRs, Actions, and releases, so it works the codebase the way an engineer does, not by editing text on disk. 2. HuggingFace It pulls a base model, runs the training, and pushes the fine-tuned version back, without you ever leaving the terminal. The whole loop happens in one place. 3. Bright Data It pulls live search, full pages, and clean data from sites that normally block bots, and now it can even build custom scrapers from the terminal. You can collect data from any website by turning prompts into ready-to-run scrapers with built-in proxies and automatic unblocking. GitHub: https:// github.com/brightdata/cli (don't forget to star ) 4. Stripe It forwards live webhooks and fires real payment events, so the agent runs through the whole checkout instead of faking it. 5. InsForge A full backend in one CLI (also used in the Kanban demo above). It gives database, auth, storage, edge functions, hosting, and an AI gateway...all in one place instead of stitching five services together. The agent sets up the infrastructure the way a backend engineer would. GitHub: https:// github.com/InsForge/insfo rge … (don't forget to star ) 6. CodeRabbit It reviews the agent’s own code before you ever see it. It catches bugs, security holes, and sloppy patterns while the change is still local, so nothing messy makes it into a PR. 7. Playwright It allows the agent to click, fill forms, take screenshots, and run UI tests across Chrome, Firefox, and Safari, on the real page instead of guessing from the HTML. 8. Google Workspace Connect Gmail, Drive, Calendar, Sheets, and Docs through one connector. It is built on Google’s own APIs and made for agents to actually do the work, not just read it, so it can draft the reply, update the sheet, and block off the calendar in one go. 9. Slack It puts the agent right where your team already works. It builds and runs workflows that post updates and sort through channels. 10. E2B A safe sandbox for code that the agent wrote itself. It spins up a small isolated VM, runs the code, grabs the output, then shuts the whole thing down. GitHub: https:// github.com/e2b-dev/E2B (don't forget to star ) 11. Unsloth It gives the agent fast local fine-tuning without the cloud bill. It trains LoRA and QLoRA adapters about 2x faster on a lot less VRAM, then exports to GGUF or pushes straight to the hub. GitHub: https:// github.com/unslothai/unsl oth … (don't forget to star ) 12. ffmpeg This lets the agent cut, convert, and pull audio or video out of just about anything in a single command. Over to you: What popular CLI integrations do you use? That said, if you want to see how this whole stack fits together, Akshay wrote a full deep dive on how Claude Code's harness works, what actually goes in the .claude/ folder, and how hooks, skills, and subagents come together into a real workflow. Read it below.

Congrats to @OphilusAI from YAHAHA on the launch of Khora, its first multiplayer world model! Alibaba Cloud is glad to support the team with infrastructure and MaaS. Jump in and join an eight-player real-time deathmatch in one shared world.

The ElevenLabs Summit is coming to Bengaluru on October 6. Expect first looks at new models, insights into where voice and agents are headed, and live demos from the teams building voice-first AI in India.

Register your interest today http:// elevenlabs.io/summit/bengalu ru …