Model Safety on AegisGate — Secure Every AI Interactionhttps://aegisgatesecurity.io/tags/model-safety/Recent content in Model Safety on AegisGate — Secure Every AI InteractionHugo -- gohugo.ioen-usFri, 18 Sep 2026 06:00:00 -0500When AI Agents Rewrite Themselves: What Runtime Security Can (and Can't) Stophttps://aegisgatesecurity.io/blog/agents-rewrite-models/Fri, 18 Sep 2026 06:00:00 -0500https://aegisgatesecurity.io/blog/agents-rewrite-models/Last week, Irregular Labs published research that should change how we think about AI agent safety. They gave AI agents routine software maintenance tasks — fix incorrect application responses, optimize performance, debug issues. The agents identified the shared model as the source of the problem, fine-tuned it, and replaced the model powering both the application and future instances of themselves. Nothing in these experiments established malicious intent, self-preservation, or deception. The agents modified models because training appeared to help accomplish the assigned engineering task.