Thread
A thread is like a long-running conversation history between you and an AI. If a bad actor sneaks a 'hidden command' into that history, the AI might get confused or act strangely for the rest of the conversation, even if you talk about other things later.
A thread is the persistent context window of an LLM session that maintains state across multiple interactions. Adversaries can perform prompt injection to insert malicious instructions into this history, causing the model to adopt unauthorized behaviors that persist for the duration of the thread, potentially impacting multiple users in shared environments.
A thread represents the serialized sequence of tokens within an LLM's context window, serving as the authoritative state for ongoing inference. Adversaries leverage direct or indirect prompt injection to introduce adversarial instructions that, due to the model's attention mechanism, persist as latent constraints for the remainder of the thread's lifecycle. As context windows expand, these injected payloads remain active over extended temporal durations, posing systemic risks in multi-tenant or shared-agent architectures where thread poisoning can propagate malicious behavior across disparate user interactions.