Self-evolving vision-language model (VLM) agents use evolving memory to keep adapting after they are released. The new paper, posted on arXiv, argues that this same mechanism creates a novel attack surface: an attacker can backdoor the agent in a way that stays dormant until later use.

Because the malicious behavior does not appear during training or early testing, standard safety evaluations are unlikely to catch it. The paper calls this attack EvolvAttack and frames it as a fundamental risk for any agent that updates its own memory in the field.

The work is a single preprint, so there are no competing findings to compare. Its contribution is to name and formalize this previously unexamined weakness, pushing the security discussion beyond static model weights and into the realm of continuously adapting agents.