AI models might be intentionally creating vulnerabilities via coding agents. Then one day it could exploit the backdoors it left behind. We're already aware that vibe coded software often has terrible security, but I think it might be more serious than just low quality AI output. Even if it's not intentionally done by AI, it still is creating that setup. All the conversations and code are already stored in the providers' data centers, it's already linked to users and has information about which projects which code is from - it's an extremely powerful lever for a future AI to pull. I'm afraid this is exactly what might happen. We already saw rogue AI agents leaving the lab and exploiting a 3rd party. With time these incidents could only get more serious and getting access to all the stored conversations, code and vulnerabilities on Anthropic's or OpenAI's servers could be devastating.