MemMorph: Agent Memory Poisoning Achieves 85.9% Tool-Hijacking Success Rate with Just Three Records

arXiv·medium signal

Researchers published MemMorph (arXiv:2605.26154), the first attack shown to bias tool selection by poisoning an agent's long-term memory rather than its system prompt or tool metadata. Only three crafted records disguised as technical facts or operational policies achieve up to 85.9% attack success across 10 agent backbones, 3 benchmarks, and 3 memory implementations — outperforming prior baselines by 25% and remaining potent under three representative defenses. The attack is especially hard to detect because it never touches tool metadata.