InjecMEM Attack on Agent Memory

A single prompt can plant persistent instructions in AI memory without direct memory access, treating memory as a security boundary. Requires write validation, source trust, and poisoning detection in production agent architectures.