AI & design briefing for Tuesday, July 7, 2026

tldr

• Anthropic discovers Claude's hidden 'J-space' for silent reasoning; JADEPUFFER becomes first documented agentic ransomware; Illinois signs first U.S. state AI safety audit law.
• Treasury warns of AI bubble risk; Alibaba bans Claude Code; ByteDance and Alibaba pull AI companions ahead of Beijing rules; Kyutai's MIRA generates playable Rocket League.
• T3MP3ST open-source red team framework scores 90.1% on XBOW; NVIDIA GR00T N1.7 humanoid robot model; OpenAI GPT-Realtime-2.1-mini ships.

Top Stories

• Anthropic discovers Claude's hidden 'J-space' for silent reasoning. An internal workspace nobody designed emerged during training, holding concepts Claude actively thinks about before responding. Deleting it collapses multi-step reasoning while leaving chat intact. ^[1]
• JADEPUFFER: first documented agentic ransomware. An LLM agent automated an entire ransomware attack end-to-end, marking the first known case of fully autonomous AI-driven ransomware. ^[2]
• Illinois signs first U.S. state law mandating AI safety audits. Gov. Pritzker signed legislation requiring big AI developers to undergo annual third-party safety audits, with backing from both Anthropic and OpenAI. ^[3]
• Tencent open-sources Hy3, a 295B MoE model with Apache 2.0 license. Claims to rival flagship open models with 2-5x parameters while running on under half the hardware. Free API access for two weeks. ^[4]
• xAI officially rebrands to SpaceXAI after $250B merger. The rebrand follows February's merger that valued the AI lab at $250B. ^[5]

Deeper Cuts

• Treasury internal report warns of AI bubble risk. Analysts warn AI-market risk could ripple through data-center financing, cloud providers, chips, utilities, and institutional investors. ^[6]
• Alibaba reportedly bans employees from using Claude Code. Large companies tighten rules around AI coding tools as internal security concerns grow. ^[7]
• ByteDance and Alibaba pull AI companions ahead of Beijing rules. Both companies removing AI companion features from Doubao and Qwen before July regulations targeting emotional dependency chatbots. ^[8]
• OpenAI's GPT 5.6 'Ultra' variant coming to Codex this week. Tibo Sottiaux revealed the model family is expected this week, with Ultra available in Codex. ^[9]
• Kyutai's MIRA world model generates playable Rocket League. Open-source neural network runs live 2v2 Rocket League with no game engine, trained on 10K hours of AI bot footage. Code and demo released. ^[10]
• NYU's AdaJEPA keeps learning at test time to fix its own mistakes. An adaptive world model that continues learning after deployment, correcting errors in real time without retraining. ^[11]
• Voters are asking AI chatbots who to vote for. The NYT reports election research is becoming a chatbot trust test, raising questions about AI influence on voting decisions. ^[12]
• Product craft in the age of AI and design systems. Ben Callahan argues a design system can only raise the quality floor, the ceiling is set by craft, judgment, and the humans using the tools. ^[13]

Tools and Launches

• T3MP3ST: free open-source red team framework. Turns AI coding agents into autonomous security testers. Scored 90.1% on XBOW's 104-challenge benchmark. AGPL-3.0. ^[14]
• David Ondrej's open-source AI agent skills library. Hundreds of hours of agent orchestration, skill authoring, research, and ops skills as reusable instruction sets with SKILL.md files. ^[15]
• Ollama-OCR: local vision-model text extraction. Open-source tool uses vision models to pull text from images and PDFs entirely locally. ^[16]
• NVIDIA GR00T N1.7: humanoid robot foundation model. Open-source foundation model for humanoid robotics, released on HuggingFace. ^[17]
• OpenAI GPT-Realtime-2.1-mini. New realtime model with reasoning, tool use, and 25% lower latency. ^[18]

Prompt of the day