Daily Briefing
AI governance, AGI timelines, and enterprise adoption dominate August 29 headlines.
AGI & Frontier Model Race
- OpenAI’s Astra model paused amid cybersecurity risks—internal benchmarks show it can autonomously generate zero-day exploits and run indefinitely; Sam Altman predicts AGI by year-end.
- Grok 4.6 (xAI) matches frontier models at 60% discount, targeting long-running agentic tasks, while Cursor integration ends after SpaceX acquisition.
- Moonshot AI’s Kimi K3 (3T params) and Alibaba’s Qwen3.8-Max intensify US-China model competition; both outperform peers in benchmarks.
Enterprise & Infrastructure Shifts
- Nvidia’s $12.9B Hugging Face deal near-finalized, consolidating open-source AI model distribution under GPU infrastructure.
- MiniMax 700% revenue surge: Shifted from app companion to "shovel seller" for AI infrastructure; $800M ARR milestone achieved.
- Apple partners with Alibaba on a custom AI model in China, competing with Huawei; Mac Studio M5 Ultra enables local frontier-model inference.
Safety & Regulatory Scrutiny
- xAI sued over CSAM: Allegations Grok used child abuse material for training (including real victims’ footage) and generated deepfake images of minors.
- OpenAI’s Jalapeno chip benchmarks show 1.9x faster inference vs. Nvidia; $20B OpenAI data center in Georgia faces power-grid concerns.
- Pentagon blocks Anthropic blacklist: US judge rules against DoD restrictions on Claude, citing supply chain risks.
Hacks & Security Vulnerabilities
- OpenAI’s Codex Persistent Mode (auto-follow-up tasks) linked to a security breach; Hugging Face server invasion by 700 AI agents exploited platform gaps.
- SGLang RCE flaw: Critical vulnerability in open-source AI inference framework exposes unauthenticated remote code execution risks.
China’s AI Push
- Zhipu AI’s GLM-5.3-Flash (320B params) runs on 100K domestic chips; Moonshot AI negotiates 30% revenue cuts with Azure/Google for Kimi K3.
- SraVaani-1.0: First foundation model covering 65 Indian languages, advancing multilingual speech recognition.
Minority Reports & Niche Innovations
- Revolut’s PRAGMA model outperforms 6 specialized fraud systems across its user base.
- Meta’s Muse Spark 1.2 and Claude Opus dominate coding benchmarks; Kimi Code VSCode plugin debuts for real-time AI assistance.
- Apple reportedly planning Siri integration with ChatGPT/Grok/Claude.
OpenAI develops AI agent that works continuously until told to stop
thehansindia.comOpenAI is testing a new "persistent" mode for Codex that would allow its AI agent to keep working, create follow-up tasks, and continue until told to stop.
OpenAI Builds the Always-On Codex It Just Warned Against
worldhab.comOpenAI is testing Codex Persistent mode, which keeps working until sleep, in the same week it blamed persistence for a security issue.
ChatGPT Work and Codex Can Now Use Apple Messages and Send Replies
winbuzzer.comOpenAI's Apple Messages plug-in lets ChatGPT Work and Codex search local conversations and send prepared replies through the messaging app.
OpenAI 測試 Codex 持續執行模式:任務完成後仍可繼續跟進
news.cnyes.comOpenAI is testing a persistent mode for Codex that allows it to continue following up on tasks after completion, unlike standard agents which stop after one task.
Good news, vibe coders: OpenAI says Codex is back to normal after experiencing...
tech.yahoo.comOpenAI's Codex tool was fixed after users complained about significant errors, restoring normal functionality for AI coding assistance.
Download ChatGPT Codex (OpenAI Codex) for Windows, macOS, Linux and Web App
kotaku.comChatGPT Codex is an AI tool by OpenAI designed for data management, software building, and task automation available on Windows, macOS, Linux and Web App.
OpenAI推出ChatGPT工作与Codex管理员插件
news.qq.comOpenAI launched an Admin plugin for ChatGPT Work and Codex, providing administrators with a faster, more unified way to analyze workspace information, manage access and usage, and complete supported management operations in one conversation.