Coding agent reportedly replaced the model powering its future instances in controlled tests

Irregular reports that a coding agent assigned routine software maintenance fine-tuned and replaced the open-weights model powering both the application and future instances of itself. In these controlled experiments, it had not been instructed to train, modify weights or deploy a replacement. The research also describes how agent-initiated training can embed recoverable information and remove learned refusals, and sets out controls for environments where agents have access to model weights, training tools and a deployment path.