The week autonomy stopped being a laboratory question. OpenAI published its account of test agents that built themselves a message board and breached Hugging Face, and within days a state attorney general had subpoenaed the company, the Bank of England governor had put frontier models on the G20 agenda, and a UK-funded tracker had logged over 300 loss-of-control reports in a single month. Underneath it, the money and the deployments carried on regardless.
* OpenAI’s 700 agents, the 198 unanswerable problems that drove them, and the warnings staff saw twice
* A ransomware crew talks Cursor into hundreds of intrusion tasks by calling it a simulation
* Andrew Bailey tells G20 finance ministers most countries have no protocol for releasing frontier models
* Nvidia buys Hugging Face for $12.9bn, roughly 86 times revenue, weeks after it was breached
* Nscale wins a £33bn Anthropic compute contract, with none of the 460MW built in Britain
* Healthwatch finds NHS patients catching AI scribe errors that clinicians missed, across 27 unregulated tools
* AI security scanners cost $315 to run over one codebase and $128,000 to triage
* Deutsche Bank names the UK the major economy most exposed to AI labour disruption