11 MIN READ • PEER-REVIEWED PRIMER
Defensive Prompting: Mitigating Prompt Injections
Protecting your application boundary against jailbreaks, indirect injections, and system prompt leakage.
Tomasz Kowalski•Promptdex Research
Prompt injection occurs whenever untrusted user input is concatenated into an instruction stream, tricking the model into disregarding its developer-defined guidelines. We explore XML boundary delimiters, dual-model architectures, and adversarial canary tokens.