Memory Poisoning

Once an AI has tools, memory and an autonomous loop, a mistake no longer stops at the answer. A tool argument, a sentence in memory, a text file in the project: an agent can promote any of them into an instruction it treats as approved.