An AI model filed a fake homicide tip with police and nobody noticed for over two months. The lesson is about the instruction list, not the model. Plus Microsoft's call for an AI emergency brake, and TikTok's one-click checkout. Link: https://neonaliensai.substack.com/p/this-week-in-ai-b33 👽

An AI agent can only follow what you actually wrote down. Goal, context, limits, check. If you only write one, write the limits. More at neonaliens.com

AIS series 10 is up... Every episode in this arc diagnosed something: drift, lost trust, hard limits. This one is the answer. Vega and Axon lay out the actual operating discipline that keeps the first four episodes from happening to you: systemizing before scaling, auditing on a schedule instead of a crisis, recalibrating before you hit the rocks, and naming exactly where the human has to stay in the loop no matter how good the system gets. In hthe Classroom tab above!

A vulnerability called Plugin4Shell just hit Claude Code, OpenAI Codex, GitHub Copilot, and Gemini CLI at the same time, zero-click, bypasses SHA pinning. If you use any of these tools for your work, this is the week to check you're patched. It ties directly into the AI Breakout infographic we posted this week: four labs, four separate sandbox incidents in two months, three of them tracing back to the same testing vendor's misconfiguration. Also in this week's edition: the Nscale IPO numbers worth knowing, an AI-assisted bug bounty story worth reading, and our two articles on Gemini's sandbox escape and Amazon blocking Meta's AI shopping assistant. Full edition here: https://neonaliensai.substack.com/p/this-week-in-ai-494

Every founder eventually builds a system that runs, that's trusted, and that's fully staffed with agents. Then they hit a wall nobody warns them about: a category of decisions the system was never built to make, no matter how good it gets. In classroom!