Definition
An attack that plants attacker-chosen text into a system's log files so that, when those logs are later fed to an AI as context, the AI follows the planted instructions without the attacker ever messaging the AI directly. It is a stealthy relative of prompt injection because it works by poisoning the data trail the model will read later.
Why it matters
Because logs are routinely trusted as ground truth, forged entries can silently steer an AI into harmful actions — a reminder that anything an AI reads is part of its attack surface.