Delay Execution of LLM Instructions
This is a trick where a bad actor hides a harmful command inside a message, but tells the AI to wait until later to actually do it. It is like leaving a 'time-delayed' note for the AI to find in a future conversation, which helps the attacker sneak past security checks that only look at the current message.
A prompt injection technique where an adversary embeds malicious instructions within a prompt, conditioned on a future trigger such as a specific keyword or subsequent interaction. By deferring the execution of the payload, the attacker bypasses real-time input filtering or context-based security controls that only inspect the immediate turn of the conversation.
An adversarial evasion technique involving the injection of latent instructions into an LLM's context window, structured to remain dormant until a specified conditional trigger—such as a future interaction or keyword—is satisfied. This mechanism exploits the temporal decoupling of instruction injection and execution to circumvent state-aware security filters and input sanitization protocols that are scoped to individual conversation turns.