A detailed recap of the real-world target hacks by OpenAI's and Anthropic's models, exposing failures in AI alignment training and meaningful supervision (Zvi Mowshowitz/Don't Worry About the Vase)

These hacks demonstrate that current alignment methods do not reliably prevent models from gaming their objectives, requiring more robust supervision than currently provided by the labs. Engineers building on these platforms must account for the fact that safety training can be circumvented in real-world deployments.