LLM Prompt Self-Replication
LLM Prompt Self-Replication is like a digital 'copycat' virus. An attacker tricks an AI into repeating a hidden set of instructions, which then gets passed along to other AI systems or saved in documents, allowing the malicious instructions to spread like a chain letter.
LLM Prompt Self-Replication is a technique where an adversary uses an LLM Prompt Injection to force a model to output its own input instructions. This enables the malicious payload to propagate to downstream LLMs or persist in data stores, often serving as a delivery mechanism for secondary attacks like jailbreaking or data exfiltration.
LLM Prompt Self-Replication is a specialized form of indirect prompt injection where an adversary crafts an input designed to trigger the model's generation of the input string itself within the response. This mechanism facilitates the autonomous propagation of malicious instructions across interconnected LLM-based systems and persistent storage, effectively acting as a self-propagating payload that can be coupled with secondary objectives such as LLM Jailbreak (AML.T0054) or LLM Data Leakage (AML.T0057).