Invisible ink used to be a child’s trick. This week it grew up and got a job in your software. A researcher showed you can hide instructions inside an AI tool in characters no screen will render, so the permission slip you read and the order the machine receives are two different documents. You said yes to the sentence you could see.
Episode two of the video experiment. So, in eight minutes: CNBC’s numbers on American companies quietly routing up to 46% of their AI traffic through Chinese open models (peak week, nearly half), the food-delivery giant that trained a 1.6-trillion-parameter model on 50,000 Chinese chips, Kai-Fu Lee selling countries their own AI stack, Rio de Janeiro’s “sovereign” model unmasked by its own weights, the invisible-ink hole in your agent’s approval screen (Invariant Labs found the class, OWASP codified it, Microsoft just warned about it), North Korea’s 90-minute npm heist, and the free school in Yerevan where teenagers flash their own Raspberry Pis and build an AI that plays rock paper scissors, in a workshop called, I’m not making this up, Owning your AI Stack.
Tell me what you think. I’m reading everything. And yes, the code behind this should be open source. I know, I know. I’m getting to it.
