This paper introduces a new attack method called adaptive long-context prompt injection (AdaLCPI) that reconstructs malicious objectives from incomplete fragments, which can be used to compromise AI agents. This highlights the need for robust safety evaluations of agents against such attacks.