alt Hacker NewsRight but the issue is users have full control over context. A security-violating action by a coding agent in one context can be completely innocuous under other contexts etc, or breaking down the task into multiple tasks that in isolation do not violate anything.
Yes, there is always a path to a problem. Even random monkeys on a keyboard can write a security exploit. Random monkeys with guidance from a knowledgeable human will do it much faster.
The goal shouldn't be to make problems impossible. It is to adjust the ratio between problems and successes.
You can also create a meta. "How much do I trust the user?" When you see the user trying to manipulate towards security, distrust the user and apply rules more strictly. If the user simply acts like a normal developer, just be a useful developer tool. Including fixing security holes when appropriate.