I think sandboxing will always be the best mitigation. Even if scanning with pre-commit hooks, the agent writing those credentials in a commit means the agent has access to them. For many use cases, it shouldn't have access from the start :) Thanks for your perspective!!