The rule is to operate using the intersection of all the users permissions of who is contributing text to the LLM. Why can an attacker's prompt access a repo th

The rule is to operate using the intersection of all the users permissions of who is contributing text to the LLM. Why can an attacker's prompt access a repo the attacker does not have access to? That's the biggest issue here.