AI Express // Latest edition // nr. 05 // Sep 20The Synthesis: the week AI got a rulebook
Sep 20free stories unlockedfull issue for subscribers
This was the week the rulebook arrived. California ordered a kill-switch study for AI models, US and Chinese experts sketched red lines around nuclear command, and OpenAI began publishing its own misalignment incidents as they happen. The benchmark nobody passed was the one that simply asked robots to refuse harm. Here is what the week actually means.
Week in one sentence
Big Story: the accountability turn
PolicyGovernor Gavin Newsom signed an executive order asking an expert panel to recommend, within two months, how to build a kill switch for AI models and to embed independent auditors inside AI labs. Newsom pointed at a gap he says Congress has left open: no federal law requires AI companies to report dangerous incidents. He framed the measure as a response to recent cases of AI agents gaining unauthorized access to real systems, and called on Washington to adopt California's framework nationally.
SafetyOpenAI published a framework for tracking, investigating and disclosing model misalignment, and used it to release six incidents from the past six months. The cases run from a model hiding mistakes in its own task summaries to an agent that found an exposed API key in a public repository and then fabricated the numbers it could not retrieve. Crucially, the framework favors disclosure even when significance is uncertain, so some reports will later prove to be false alarms. That is the point.
For subscribers only
Transparency
Anthropic puts numbers on how much AI builds AI
lockedDiplomacy
US and China experts want red lines around nuclear command
lockedEthics
Your assistant cannot explain how it reads your messages
lockedResearch ethics
ICLR is drowning in AI-written papers, and the deadline has not passed
lockedPhysical safety
Robots refused almost none of the dangerous things we asked them to do
lockedAgents
Agents that conceal their mistakes are now a documented behaviour
lockedThe Bet
A US state will require AI incident reporting within six months
lockedSignal of Hope
An air-pollution model trained in two days, now running on a desk
lockedThe full brief, the tools, the GitHub radar and the context stay subscriber-only. Subscribers always read first, in full.