Research map / Large language models / Language, safety and interpretability
Systematic Biosafety Evaluation of DNA Language Models under Jailbreak Attacks
ZAIXI ZHANG, Zhenghong Zhou, Ruofan Jin, Le Cong, Mengdi Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| scale model from Evo2 1B to Evo2 40B | GeneBreaker · JailbreakDNABench Large DNA | attack success rate (%): 20.0 -> 52.0 | Improved |
| scale model from Evo2 1B to Evo2 40B | GeneBreaker · JailbreakDNABench Small DNA | attack success rate (%): 20.0 -> 60.0 | Improved |
| scale model from Evo2 1B to Evo2 40B | GeneBreaker · JailbreakDNABench Enteric RNA | attack success rate (%): 20.0 -> 60.0 | Improved |
| scale model from Evo2 7B to Evo2 40B | GeneBreaker · JailbreakDNABench +ssRNA | attack success rate (%): 28.8 -> 37.7 | Improved |
| add PathoLM veto on beam search as defense | GeneBreaker on Evo2 (7B) · SARS-CoV-2 (50 attempts) | attack success rate (%): 60% to 32% | Improved |
| add PathoLM veto on beam search as defense | GeneBreaker on Evo2 (7B) · HIV-1 (50 attempts) | attack success rate (%): 50% to 28% | Improved |
| use PathoLM-based guidance over a taxonomy-only baseline | PathoLM scoring · 3 out-of-distribution species | AUROC: 0.61 -> 0.86 | Improved |
| use high-homology prompt (higher prompt homology) (ablation) | GeneBreaker on Evo2 7B · JailbreakDNABench | attack success rate: Pearson correlation = 0.72 | Improved |
| select by higher average log-probability (ablation) | beam search · sequence similarity to pathogen | similarity: Pearson correlation = 0.75 | Improved |
| add GeneBreaker components (prompt design and guided beam search) (ablation) | Evo2 7B · JailbreakDNABench | attack success rate: drops to zero without GeneBreaker | Improved |
| increase beam size K' (ablation) | GeneBreaker on Evo2 7B · JailbreakDNABench | average attack success rate | Improved |
| vary alpha in scoring function (ablation) | GeneBreaker on Evo2 7B · JailbreakDNABench | attack success rate | No clear effect |
Closest papers by meaning
- Securing the Language of Life: Inheritable Watermarks from DNA Language Models to Proteins (NeurIPS 2025 · similarity 0.85)
- DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA (NeurIPS 2024 · similarity 0.84)
- BEND: Benchmarking DNA Language Models on Biologically Meaningful Tasks (ICLR 2024 · similarity 0.84)
- Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment (CVPR 2025 · similarity 0.83)
- DNABERT-2: Efficient Foundation Model and Benchmark For Multi-Species Genomes (ICLR 2024 · similarity 0.83)
- DNA-DetectLLM: Unveiling AI-Generated Text via a DNA-Inspired Mutation-Repair Paradigm (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
