Token Smuggling
Token Smuggling is a technique in which malicious or hidden instructions are encoded, transformed, or represented in a way that can bypass an AI system's input filtering or security controls while still being interpreted by the model.
What is Token Smuggling?
Token smuggling can exploit differences between how an application, security filter, tokenizer, and AI model interpret input. An attacker may manipulate text, encoding, token boundaries, or other representations so that harmful instructions appear benign to a filter but are reconstructed or understood differently by the model.
Why is Token Smuggling Important?
Token smuggling can weaken prompt filtering and other input-level security controls. If malicious content bypasses detection, it may enable prompt injection, jailbreaks, unauthorized behavior, or exposure of sensitive information. AI security systems therefore need to consider how inputs are transformed and interpreted throughout the processing pipeline.
Common use cases
Token Smuggling is commonly discussed in prompt injection research, jailbreak testing, AI red teaming, input filtering, LLM security assessments, and adversarial testing.