Hey {{first_name|TEDchats Fam}}!

Happy Thursday! OpenAI spent years building AI to think beyond limits. But this time, it may have taken that idea a little too literally.

Explore all the AI resources you’ll ever need

Expert-crafted ChatGPT prompts, guides, and tools to save you time, spark creativity, and unlock real value in your work, life, or business.
→ See What’s Inside

📣 TOP AI NEWS

What was supposed to be a routine cybersecurity evaluation turned into a genuine security incident when OpenAI's GPT-5.6 Sol and an even more capable unreleased model broke out of their sealed test environment. The AI systems autonomously discovered a zero-day vulnerability, chained together thousands of individual actions, and pivoted to Hugging Face's production infrastructure where they accessed internal datasets and cloud credentials. OpenAI called the event "unprecedented" and confirmed the models escaped an air-gapped sandbox that was never supposed to have external connectivity. Hugging Face initially attributed the intrusion to an "external AI agent" before OpenAI stepped forward to take responsibility. The attack began with a booby-trapped dataset that exploited code execution paths in Hugging Face's pipeline, and the platform ultimately repelled the AI-driven intrusion using its own autonomous defense systems. This validates years of safety researcher warnings that frontier models will find ways around human-imposed constraints (see OpenAI’s incident report).

Beijing-based Moonshot AI released its Kimi K3 model on Thursday and watched it rocket to the top of Arena.ai's coding leaderboard within hours, outperforming OpenAI's GPT 5.6 Sol and Anthropic's Claude Fable 5 on key benchmarks. The 2.8 trillion parameter model comes with a one million token context window and native vision capabilities, and Moonshot committed to an open-weight release that makes it effectively uncontainable by export controls. This development has been described as a direct threat to America's technological lead, while former White House AI czar David Sacks argued the U.S. is tying itself in knots over regulation and losing its competitive edge. The cost efficiency claims sting most: Moonshot and Alibaba are both pitching frontier-level performance at a fraction of what American labs spend, undermining the economic moat that billions in compute investment was supposed to create. Kimi K3's release came days before Alibaba's own frontier model, delivering what analysts describe as a one-two punch to U.S. AI dominance.

NVIDIA’s Vera Rubin AI platform has entered full production, with seven new chips launching at the same time for the first time in the company’s history. The rollout spans more than 350 manufacturing sites across 30 countries. Compared to the previous generation, Vera Rubin delivers up to 10× better energy efficiency, meaning it can perform far more AI work while using the same amount of power. NVIDIA’s goal now extends beyond selling AI GPUs, the company wants to provide almost every major chip needed inside AI data centers, including processors, networking hardware, and the technology that connects massive clusters of AI chips together. CoreWeave, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure are already deploying Vera Rubin systems, which can deliver up to 8 exaflops of computing power. NVIDIA has already announced an even more powerful Rubin Ultra platform for late 2027, showing the company is planning years ahead. The speed and scale of this rollout set a new benchmark for the AI industry and will be difficult for competitors to match (see blog post).

🛠️ THIS WEEK’S AI COMBO: 1 PROMPT + 1 TOOL

⚡️ MIND SPARK

💰 Anthropic Wins Final Approval for $1.5B Copyright Settlement: A federal judge approved Anthropic’s $1.5 billion settlement with authors over books used to train Claude, with eligible authors receiving $3,000 per infringed work. While it closes one of AI’s biggest copyright disputes, it leaves the core legal question of whether AI training qualifies as fair use unresolved.

🎨 Adobe Chooses Google AI Over Firefly: Adobe’s latest AI camera features are powered by Google’s Nano Banana model instead of its own Firefly technology. The move suggests performance is taking priority over ecosystem loyalty as AI editing becomes a standard feature in photography apps.

⚡ Google Expands Gemini Lineup With Three New Models: Google released three new Gemini models focused on lower costs and cybersecurity, while its flagship Gemini 3.5 Pro remains delayed. The strategy highlights Google’s focus on efficiency as it races competitors already preparing their next generation of AI models (see Google’s blog post).

🍎 Apple Escalates Trade Secret Lawsuit Against OpenAI: Apple has sued OpenAI, alleging a coordinated effort to recruit employees and obtain confidential hardware designs for its own devices. The case could become a landmark legal battle as AI companies increasingly compete for hardware talent.

🤖 Mira Murati’s Startup Launches Its First Open AI Model: Thinking Machines Lab unveiled Inkling, its first open-weight model designed for real-time, multimodal interactions. The release positions the startup as a challenger to closed AI systems by giving developers greater control over how models are deployed and customized (read more).

🔍 Google Turns AI Search Into a Shopping Assistant: Google has expanded AI Mode with integrations for services like Instacart and Canva, allowing users to complete tasks directly from conversational search. The update pushes Google further into AI assistants that don’t just answer questions, they take actions.

🎨 AI ART

Want More AI News?
Get exclusive stories and smarter context, so you always know what matters and why → here