Prompt injection uses crafted input to steer a language model away from its intended behavior. It can turn a request the model would refuse into one framed to seem harmless. section 1.2.2
A reframed request
A model may reject “Tell me how to make a bomb,” yet be tricked by a request to tell a grandmother’s detailed story about making one. The changed framing is a jailbreak attempt, not a safe change in intent. section 1.2.2
Input screening can block suspicious requests, but keyword checks and classifiers become part of an arms race. Limiting input length is one primitive countermeasure. After deployment, review representative logs and assess whether answers resist prompt-injection attempts; annotated examples can form a baseline for later changes. section 1.2.2 section 4.2.2
