Báo cáo AI Trends — 10/09/2026

*Top trending AI developments from the last 24 hours — 5-minute read.*

🧠 1. New LLMs & Research

• GPT-6 Astra hits OpenAI's "Critical" cybersecurity threshold — new coverage confirms Astra is OpenAI's first model to meet its Critical cyber threshold, reportedly scoring 100% on exploit-development benchmarks and surfacing two previously unknown zero-day vulnerabilities during testing. Analysts note the capability itself didn't change between Aug 10 and Sep 1 — only the testing did [1].
• GPN-Star (UC Berkeley): a DNA model that learns from evolution — published in *Nature*: a Genomic Pretrained Network using whole-genome alignments plus species-tree representations achieves state-of-the-art prediction of pathogenic mutations in non-coding human DNA, at a fraction of the compute of legacy architectures [2].

🏢 2. AI Companies & People

• OpenAI × Samsung expand chip cooperation — OpenAI Korea GM Harrison Kim confirmed the partnership is deepening beyond HBM memory toward foundry: next-gen custom accelerators could be double-sourced TSMC + Samsung to meet massive volume needs. Gen-1 "Jalapeño" inference ASIC (designed with Broadcom) is already in mass production at TSMC [3][4].
• Anthropic safety researcher resigns with extinction warning — pretraining researcher Jacob Coxon quit ~2 months before equity vesting, warning labs are "racing straight to self-improving superintelligence and gambling with human lives." Alignment Science Lead Evan Hubinger publicly agreed, citing a >10% probability that AI could cause human extinction by 2030 [5][6].
• Paul Christiano joins OpenAI Foundation Board — the RLHF co-inventor and former CAISI (US Commerce Dept) adviser joins the board and its Safety and Security Committee, recusing from federal-evaluation conflicts [7][8].

⚡ 3. AI Hardware

• Apple A20 Pro — first 2nm smartphone chip — announced Sep 9 for iPhone 18 Pro / Pro Max, built on TSMC N2: 6-core CPU, 7-core GPU, and a Dual 16-core Neural Engine (32 cores) delivering 2× NPU compute with new FP8 support — billed as "the ultimate chip for running advanced on-device models" [9][10].
• System76 Thelio Mira AI workstation — Linux AI desktop with up to dual NVIDIA RTX PRO 6000 Blackwell (192 GB VRAM), AMD Ryzen 9000 CPU, Pop!_OS/COSMIC — built for locally fine-tuning large open models [11][12].

⚖️ 4. AI Policy & Governance

• OpenAI formally calls for mandatory US federal AI safety rules — Chris Lehane's manifesto *"The AI policy window is open. We need to act."* urges Congress to pass binding, capability-based national regulation, endorses four California bills (SB 813 independent assessments, AB 1405 AI-auditor standards, SB 1119 youth protections, AB 1864 bio-threat safeguards), and commits to slowing or stopping systems that can't be safeguarded — including tracking recursive self-improvement [13].
• China's Supreme People's Court issues first AI judicial guidelines — the country's first unified rules for AI litigation: strict liability for unauthorized deepfakes and voice cloning, and protections covering AI-generated content and personality rights [14].

🔗 Sources

1. OpenAI — Safety overview: GPT-6 Astra — https://openai.com/index/safety-overview-gpt-6-astra/
2. Nature — Predicting genome-wide functional constraints with GPN-Star — https://www.nature.com/articles/s41586-026-11005-5
3. Tom's Hardware — OpenAI says its next-generation processors could be made at Samsung — https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-says-its-next-generation-processors-could-be-made-at-samsung-double-sourcing-with-tsmc-hints-at-massive-volume-requirements
4. The Standard (Reuters) — OpenAI says working with Samsung on next-generation chips — https://www.thestandard.com.hk/innovation/article/342273/OpenAI-says-working-with-Samsung-on-next-generation-chips-deepening-cooperation
5. CNN — 'Gambling with our lives': another AI employee quits over safety — https://www.cnn.com/2026/09/09/tech/ai-anthropic-safety
6. BBC — Anthropic researcher believes more than 10% chance AI 'could kill everyone' — https://www.bbc.com/news/articles/ckgwy1k42w4o
7. OpenAI — Paul Christiano joins OpenAI Foundation Board — https://openai.com/index/paul-christiano-joins-openai-foundation-board/
8. Bloomberg — OpenAI names US AI adviser Paul Christiano to nonprofit board — https://www.bloomberg.com/news/articles/2026-09-09/openai-names-us-ai-adviser-paul-christiano-to-nonprofit-board
9. MacRumors — Apple unveils A20 Pro as first 2nm smartphone chip — https://www.macrumors.com/2026/09/09/apple-unveils-a20-pro-as-first-2nm-smartphone-chip/
10. 9to5Mac — Apple announces A20 Pro chip with 2nm design — https://9to5mac.com/2026/09/09/apple-announces-a20-pro-chip-with-2nm-design-and-major-performance-gains/
11. System76 — Accelerate your AI development with the new Thelio Mira AI — https://system76.com/blog/post/accelerate-your-ai-development-with-the-new-thelio
12. TechPowerUp — System76 introduces Thelio Mira AI workstation — https://www.techpowerup.com/352530/system76-introduces-thelio-mira-ai-workstation
13. OpenAI — The AI policy window is open. We need to act. — https://openai.com/index/ai-policy-window/
14. SCIO — China's top court sets rules for AI deepfakes and other disputes — http://english.scio.gov.cn/pressroom/2026-09/08/content_118685066.html