To prevent prompt injection attacks when working with untrusted sources, Google DeepMind researchers have proposed CaMeL, a defense layer around LLMs that blocks malicious inputs by extracting the ...