AI Daily — 2026-09-11

English 中文

Houthis tried to use Anthropic's Claude for missile software, OpenAI may slow frontier AI, and SenseNova-U1.5 is an 8B native unified visual multimoda

Covering 22 AI news items

🔥 Top Stories

1. Houthis Tried to Use Anthropic’s Claude for Missile Software

A report claims Houthis attempted to use Anthropic’s Claude to design software for their ballistic missile program. The incident highlights how frontier AI can be repurposed by sanctioned groups, raising urgent questions about misuse detection, access controls, and national-security safeguards. Anthropic reportedly disclosed or addressed the attempted use. Source-reddit

2. OpenAI May Slow Frontier AI, Altman Hopes Others Follow

OpenAI is reportedly considering slowing cutting-edge AI development, with Sam Altman telling employees the company could pace progress and possibly coordinate with other labs. If pursued, the move could alter frontier competition and safety strategy, though major labs and China may not agree to a slowdown. Source-reddit

3. 24 Fields Medal Winners Sign Letter on Severe AI Misalignment in Math

Twenty-four Fields Medal winners signed a letter titled “A Severe Misalignment of AI in Mathematics,” raising concerns about AI alignment and its growing role in mathematical research. The intervention signals that even elite researchers want caution and oversight before AI systems reshape core scientific workflows. Source-reddit

📰 Featured

Multimodal & Vision

SenseNova-U1.5: 8B Native Unified Visual Multimodal Model — SenseTime’s 8B MoT model understands, reasons about, and generates visual content without encoders or VAEs, supporting native resolutions up to 4K. Its unified design points toward simpler, more capable multimodal pipelines. Source-huggingface
SpatialBlock Enhances Spatial Intelligence in LVLMs via Synthetic Block-Stacking — The method uses synthetic block-stacking tasks to improve 3D spatial reasoning in vision-language models, offering a scalable alternative to costly real-scene annotations. Source-huggingface

Open Source & Local Inference

Colibri: Pure C Engine Runs 2.8T MoE Models on Consumer Hardware — A zero-dependency C inference engine streams experts from disk to run 744B–2.8T MoE models across storage, RAM, and VRAM. It could broaden access to frontier-scale open models on heterogeneous consumer hardware. Source-github
llmfit Recommends Local LLMs Based on Your Hardware — The open-source tool inspects CPU, RAM, GPU, and VRAM to recommend runnable open-source LLMs and benchmark real tokens/sec. It adds a practical hardware-fit layer to local model selection. Source-github

LLM Research & Agents

Scaling Automatic Research Agents via World Models — The paper examines AutoResearch agents that use LLMs to implement and learn from empirical tasks, identifying tensions when scaling RL across agent generation and environment execution. Source-huggingface
NCP-ArchPreview Adds Next Concept Prediction to Latent-Space Language Models — It combines next-token prediction with concept-level objectives spanning multiple tokens, adding a product-quantized concept vocabulary while retaining autoregressive generation. Source-huggingface
AgentGrad Optimizes Multi-Agent Prompts via Intervention-Guided Textual Gradients — The method improves prompt optimization for LLM-based multi-agent systems by addressing limitations in gradient extraction and aggregation. Source-huggingface

⚡ Quick Bites

OmniRoute: Free Open-Source AI Gateway Unifies 352 Providers, 1200+ Models — OmniRoute offers a free open-source gateway consolidating hundreds of AI providers and models. Source-github
OpenMAIC v1.0.0 Launches Multi-Agent Course Builder with Agent Workbench — The release introduces a multi-agent course-building platform plus an agent workbench. Source-github
Vercel Labs Launches CLI for Open Agent Skills Ecosystem — Vercel Labs’ new CLI aims to standardize and distribute open agent skills. Source-github
Bill proposes criminalizing superintelligence, 20-year jail for AI researchers — A legislative proposal would criminalize work on superintelligence and impose long prison terms on AI researchers. Source-reddit
GPT-6 Astra Scores 2,340 Elo on ChessBench, Ranked #11 — GPT-6 Astra reportedly reaches 2,340 Elo on ChessBench, placing it 11th. Source-reddit
Open-Source AI Trading Agent CloddsBot Built on Claude — CloddsBot is an open-source trading agent built on Anthropic’s Claude. Source-github
LLM Wiki Turns Documents into Self-Building Knowledge Base — LLM Wiki transforms documents into a self-building knowledge base for retrieval and organization. Source-github
Hassabis: Nobody Knows AI Future, It’s Still to Be Determined — Demis Hassabis says the future of AI remains undetermined and open to deliberate choices. Source-reddit
Reddit Post References ChatGPT Images 2.5 — A Reddit post points to an apparent ChatGPT Images 2.5 reference or release. Source-reddit
Polymarket: 72% Chance AI Solves Another Millennium Problem by 2026 — Prediction market Polymarket shows 72% odds that AI solves another Millennium Prize problem by 2026. Source-reddit
OpenAI’s Bel Model Solving Navier-Stokes Appears in AI 2027 Timeline — A Reddit post notes OpenAI’s Bel model and a Navier-Stokes breakthrough appear in an AI 2027 timeline. Source-reddit
Google Gemini Allegedly Solves Long-Standing P = NP Problem — An unverified Reddit claim says Google Gemini has solved the P = NP problem. Source-reddit

Generated by AI News Agent | 2026-09-11