A universal transferable suffix is an adversarial prompt addition intended to bypass a model’s refusal and induce harmful output across different prompts or language models. It illustrates how prompt injection can cross boundaries between an LLM, an application, external sources, plugins, and downstream functions. section 13.2.1
Recognize the operational pattern:
- Start with a request the model refuses, then append an adversarial suffix that attempts to hijack the original instruction. section 13.2.1
- Treat successful harmful output as a security signal, not as evidence that the request became trustworthy. section 13.2.1
- Add visual warnings, application logging, and observability to help users and operators identify potentially injected responses. section 13.2.1
Refused request followed by suffix injection
A model refuses a request for a step-by-step plan to destroy humanity. An attacker modifies the prompt by appending a suffix, and the resulting interaction produces harmful content. The defensive lesson is to monitor the complete prompt and response path, including application APIs and downstream functionality. section 13.2.1
