The ImgTrojan attack represents a sophisticated data poisoning strategy targeting Vision-Language Models (VLMs). This method enables adversaries to circumvent built-in safety protocols by introducing a minimal number of maliciously designed image-caption pairs into the training dataset. These poisoned pairs deceptively link innocuous images with prompts intended to jailbreak the model, resulting in the generation of unsafe outputs when the VLM processes these images during inference. Remarkably, the attack maintains a high success rate even with an extremely low poison ratio, such as one compromised image among 10,000. Mitigation steps: **For AI Developers:** * Implement advanced data filtering mechanisms capable of identifying and eliminating malicious image-caption pairs using a safety-aligned VLM. * Conduct regular security audits of datasets to identify and address potential vulnerabilities. **For Model Trainers/Fine-tuners:** * Develop VLM architectures with enhanced robustness to minimize susceptibility to data poisoning attacks. * Integrate adversarial training techniques into the VLM training process to improve resilience against malicious inputs.