Today’s top line: US Congress has summoned Nvidia’s CEO amid Chinese espionage concerns, researchers found 800+ publicly exposed Artex AI hacking deployments, and Google DeepMind cut Gemini video prices while extending clip length to 15 seconds.
US Congress summons Nvidia CEO over China espionage concerns
Summary: The chairs of three House committees asked Nvidia CEO Jensen Huang to appear at a closed-door briefing on November 18 to discuss reports about Chinese access to sensitive US technology. The New York Times reported US intelligence believed Huang had been targeted in a Chinese espionage operation involving an executive at Megaspeed.
Digest: A congressional briefing raises new scrutiny of Nvidia’s customer relationships and supply-chain security at a moment when US export controls and chip smuggling investigations are already intensifying. Expect further oversight and possible follow-ups after the briefing.
Source: Ars Technica
Researchers find 800+ publicly exposed Artex AI hacking deployments
Summary: Security firm Censys reported it found more than 800 distinct Artex deployments accessible from the web without authentication. Artex is an open-source autonomous penetration-testing tool linked to multiple attacks since its July release; Microsoft and CrowdStrike have tied Artex to high‑speed cloud intrusions against customers.
Digest: Misconfiguration is a major risk vector: powerful attacker tooling exposed online lowers the bar for large-scale breaches and complicates defence. Organisations should prioritise configuration audits, multifactor authentication and monitoring for agentic attack tooling.
Source: TechCrunch
Google DeepMind cuts Gemini Omni video price and extends clip length
Summary: DeepMind rolled out Gemini Omni 1.2 Flash across the Gemini API, Google Vids and Gemini for Business. It adds native 15-second video creation (up from 10 seconds), multi-turn conversational editing and stronger reference adherence, and lowers price by 30% to $0.07 per second for 720p video.
Digest: Lower costs, longer clips and editing workflows make AI video generation more viable for advertising, product clips and scale use. Expect faster adoption in marketing and content pipelines and new scrutiny over brand and IP consistency.
Source: Google DeepMind
Blue-wave midterms could reshape US AI industry and data centres
Summary: TechCrunch reports Democrats’ strong midterm results coincided with 41 ballot measures touching data‑centre development, where voters in all but one case backed more restrictions. Newly elected officials are expected to increase scrutiny of Big Tech and data‑centre expansion.
Digest: Expect a tougher regulatory environment for AI infrastructure and higher local political risk for new data‑centre projects. Companies should reassess site-selection, community engagement and regulatory scenarios for capacity planning.
Source: TechCrunch
BlockLock paper shows bitwise reproducible LLM inference with big speed gains
Summary: An arXiv paper presents BlockLock, a method to make LLM inference reproducible at the bit level while retaining batched compute. The authors evaluate five dense models and three single‑GPU deployments, and report that BlockLock-Rows cut latency by about 53.5–55.0% and raised throughput by 115.2–122.2% versus vLLM Batch‑Invariant.
Digest: Reproducible outputs matter for compliance, debugging and safety. BlockLock promises a practical path to deterministic inference without the full overhead of batch‑invariant kernels, which could influence production deployment designs.
Source: arXiv
KDASO paper automates evolving skill libraries for liability adjudication
Summary: The KDASO arXiv paper proposes a Knowledge‑Data Dual‑Driven Automated Skill Optimization system that continuously updates executable skills for liability adjudication tasks. The authors report KDASO outperforms baselines across ride‑hailing adjudication, legal liability and logical reasoning tasks and yields an 8.96% improvement when transferring skills to similar scenarios.
Digest: For businesses using LLMs in regulated decision workflows, KDASO points to automated methods for keeping rule‑like skills current and aligned with real cases – lowering manual maintenance and improving transferability.
Source: arXiv
NVIDIA releases Cosmos 3 Nano open models for interactive world simulation
Summary: NVIDIA published four Cosmos 3 Nano post‑trained models (16B trainable parameters) on Hugging Face under OpenMDW1.1. The checkpoints support multi‑sensor simulation: synchronized video and LiDAR generation, scene and traffic controls, streaming generation, and tools for autonomous‑driving and robotics research.
Digest: Open checkpoints and training recipes lower the barrier to building richer simulated worlds for policy learning and evaluation. Companies building autonomous systems and simulators can adapt these models to expand training data and test closed‑loop behaviours.
Source: Hugging Face
Quick hits
- Meta launches Spaces for Muse agents – Meta added Spaces, topic-focused collaboration tabs for Muse agents and integrates third‑party apps to keep trackers and planners updated (TechCrunch).
- OpenAI partners with Hacktron to test system security – OpenAI announced a partnership to find vulnerabilities and test third‑party connections to strengthen defences (OpenAI).
- ToyBlocks framework for robot learning and evaluation – Hugging Face and AllenAI released ToyBlocks, a controlled simulation world that auto‑generates tasks and demonstrations for robot learning studies (Hugging Face).
- Google clarifies school AI policy – search still AI-first – Google says schools can opt out of Gemini apps but cannot use Google Search without its AI features; schools may disable search entirely if they want no AI (Ars Technica).