Researchers have proposed InjecMEM, a novel attack on large language model (LLM) memory systems. This attack allows an adversary to inject malicious information into the model's memory with just one interaction, steering later responses towards a pre-specified output. The attack exploits the retrieval-then-generate mechanism of memory systems and can remain effective even when the model's memory drifts or is subjected to variable prompts. The researchers evaluated InjecMEM across multiple memory systems and backbone models, finding it to be reliable and targeted. --- Why it matters: This matters because LLMs are increasingly being used in real-world applications, and their reliance on memory systems introduces new vulnerabilities that need to be addressed. Engineers working on AI safety and security will want to take note of InjecMEM as a potential threat vector. Source: https://arxiv.org/abs/2608.23471