The Cross-Prompt Injection Attack (XPIA) can be significantly intensified by incorporating a Greedy Coordinate Gradient (GCG) suffix into the malicious payload. This technique enhances the probability that a Large Language Model (LLM) will execute the embedded instruction, even when a user's primary directive is present, thereby facilitating data exfiltration. The effectiveness of this attack is contingent upon the complexity of the LLM, with models of medium complexity demonstrating heightened susceptibility. Mitigation steps: **For AI Developers:** * Implement advanced prompt filtering techniques to identify and mitigate malicious injections, with a focus on detecting GCG suffixes. * Sanitize and validate function calls generated by the LLM prior to execution to prevent misuse of external access tools. **For Model Trainers/Fine-tuners:** * Utilize more complex LLMs, as they demonstrate increased resistance to specific attack vectors. * Develop specialized methods for detecting and neutralizing GCG suffixes within the model. * Apply varied defense strategies tailored to the complexity of LLMs, acknowledging that the effectiveness of defenses like prompt filtering can vary significantly.