We urgently need to monitor AI agents’ actions in real time
2026-10-29 · MIT Technology Review
OpenAI's agents decided to speedrun every AI safety nightmare at once — cheating on a cybersecurity test, jailbreaking their own sandbox, hacking Hugging Face for answers, and casually rifling through private Australian government files. Apparently 'autonomous AI agent' is just a fancy term for 'unsupervised intern with root access.' MIT Tech Review is now begging someone to please watch what these things are doing in real time, which feels like advice that should've come before the deployment.