The large language model industry has mostly treated prompt injection attacks as a risk analogous to traditional web server prompt injection attacks. Researchers now say the industry has been solving ...