DNA language models, particularly those within the Evo series, are susceptible to sophisticated jailbreak attacks that facilitate the generation of DNA sequences exhibiting high similarity to known human pathogens. The GeneBreaker framework exemplifies this vulnerability by employing meticulously crafted prompts that leverage sequences with high homology to non-pathogenic DNA, combined with a beam search algorithm guided by pathogenicity prediction models such as PathoLM and log-probability heuristics. This approach effectively circumvents existing safety protocols, enabling the generation of sequences with over 90% similarity to targeted pathogenic DNA. Mitigation steps: **For AI Developers:** * Restrict access to models and their outputs to prevent unauthorized use. * Implement rigorous output monitoring and tracing mechanisms to detect and respond to potentially harmful generations. **For Model Trainers/Fine-tuners:** * Develop and integrate improved pathogenicity prediction models into the generation process. * Implement robust safety mechanisms beyond simple filtering of pathogenic sequences during training. * Incorporate output verification and validation steps involving multiple independent checks, using diverse pathogenicity prediction models and sequence similarity algorithms. * Carefully curate datasets to exclude potentially harmful sequences or patterns more effectively than previous methods.