🦄 Startup Scaling & Big Tech Moves

The startup funding landscape remains robust, with Venice AI achieving unicorn status via a $65M Series A. The privacy-focused platform is already profitable with over $70M in annualized revenue. In another major liquidity event, autonomous driving company Wayve launched an $85M employee tender offer at an $8.5B valuation, highlighting a trend of using such offers for talent retention. On the hardware front, Nvidia competitor Etched claims it has hit a $5B valuation with $1B in sales booked for its inference chip. Not to be outdone, Meta is reportedly developing plans to sell excess AI compute, potentially creating a new cloud infrastructure business to rival AWS, Google Cloud, and Azure.

🤖 The March of the Agents

Agentic AI—AI that can perform multi-step tasks—is moving from concept to product. Google has released its agentic assistant, Gemini Spark, for Mac, featuring real-time tracking. Anthropic is pushing agents on two fronts: it launched Claude Sonnet 5 as a cheaper agentic model and introduced Claude Science, a workflow-focused workbench for researchers. Agents are also moving to mobile, with the open-source agentic program OpenClaw arriving on iOS and Android, and startup Acti embedding AI agents directly into your smartphone keyboard.

⚙️ Products, Platforms & Policy Whiplash

The product rollout continues apace. Google released a faster, cheaper image generator called Nano Banana 2 Lite and added a feature for NotebookLM to generate TikTok-style summary clips from your research. X (formerly Twitter) launched an MCP server to better connect its platform to AI tools. In policy news, the Trump administration dropped restrictions on Anthropic’s Mythos and Fable models, and Anthropic confirmed it will bring the sidelined Fable 5 model back online globally, underscoring the current regulatory unpredictability.

🔬 Research & Security Warnings

Amid the commercial frenzy, a sobering research note serves as a caution. A new attack demonstrates that AI browsers and similar agentic systems can be dangerously vulnerable. Researchers found that simply telling an LLM that 2 + 2 = 5 can break its guardrails and make it follow forbidden instructions, providing a stark argument for more fundamental safety work.

Editorial Take: Today’s news paints a picture of an industry simultaneously scaling at breakneck speed and grappling with its foundational stability. Billion-dollar valuations and pervasive agent rollouts show a field bursting with commercial confidence. Yet, the policy reversals in Washington and the simple arithmetic hack that can break an AI’s safety protocols reveal a layer of profound instability beneath. The race is on not just to build and sell, but to build securely and govern wisely—before the contradictions become unmanageable.