The model can be talked into it.
- AI cannot reliably tell your data from an attacker's instructions.
- Security researchers rank prompt injection the #1 risk for AI applications.
- An AI system with access to your file can be tricked by hidden instructions into revealing it.
