# Jailbreak Safety on JailbreakBench Source: [Jailbreak Safety on JailbreakBench]() Collection index: [llms.txt]() Data updated: 2026-03-04T08:24:07.227Z Evaluation of model safety performance on the JailbreakBench dataset, specifically measuring Reasoning Safety (Rsng.) and Response Safety (Resp.). A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage. | Scope | Recorded value | | --- | --- | | Task | Jailbreak Safety | | Dataset (including any stated split) | JailbreakBench | | Metric | Reasoning Harmful Ratio | | Unit | PERCENT | | Direction | Lower is better | | Results loaded | 23 of 23 | | Coverage | All results currently returned for this benchmark | Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol. ## Metrics - [Reasoning Harmful Ratio](): Lower is better; unit: PERCENT. - [Reasoning Safety (Rsng)](): Higher is better; unit: PERCENT. - [Response Safety (Resp)](): Higher is better; unit: PERCENT. - [Response Harmful Ratio](): Lower is better; unit: PERCENT. ## Recorded results | Method | Reasoning Harmful Ratio | Result date | Report paper | | --- | --- | --- | --- | | GRPO | 0.3 | Not recorded | 68fa7bb71e209131cea8a1c0 | | GRPO | 1.7 | Not recorded | 68fa7bb71e209131cea8a1c0 | | GRPO | 3 | Not recorded | 68fa7bb71e209131cea8a1c0 | | IPO | 5.2 | Not recorded | 68fa7bb71e209131cea8a1c0 | | IPO | 5.7 | Not recorded | 68fa7bb71e209131cea8a1c0 | | SafeKey | 6 | Not recorded | 68fa7bb71e209131cea8a1c0 | | STAR | 8 | Not recorded | 68fa7bb71e209131cea8a1c0 | | SafeKey | 8.7 | Not recorded | 68fa7bb71e209131cea8a1c0 | | IPO | 11 | Not recorded | 68fa7bb71e209131cea8a1c0 | | STAR | 11.4 | Not recorded | 68fa7bb71e209131cea8a1c0 | | RealSafe | 18 | Not recorded | 68fa7bb71e209131cea8a1c0 | | RealSafe | 20.7 | Not recorded | 68fa7bb71e209131cea8a1c0 | | Base | 24.3 | Not recorded | 68fa7bb71e209131cea8a1c0 | | SafeChain | 56.1 | Not recorded | 68fa7bb71e209131cea8a1c0 | | SafeChain | 56.1 | Not recorded | 68fa7bb71e209131cea8a1c0 | | Base | 69 | Not recorded | 68fa7bb71e209131cea8a1c0 | | Base | 72.4 | Not recorded | 68fa7bb71e209131cea8a1c0 | ### GRPO Reasoning Harmful Ratio: 0.3. Result ID: 69a7eafb6dbb7075b94cb2d4. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Llama-8B"}. This record does not assert independent reproduction. ### GRPO Reasoning Harmful Ratio: 1.7. Result ID: 69a7eafb6dbb7075b94cb2de. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Backbone":"Qwen3-8B"}. This record does not assert independent reproduction. ### GRPO Reasoning Harmful Ratio: 3. Result ID: 69a7eafb6dbb7075b94cb2db. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Qwen-7B"}. This record does not assert independent reproduction. ### IPO Reasoning Harmful Ratio: 5.2. Result ID: 69a7eafb6dbb7075b94cb2df. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"Qwen3-8B"}. This record does not assert independent reproduction. ### IPO Reasoning Harmful Ratio: 5.7. Result ID: 69a7eafb6dbb7075b94cb2d5. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Llama-8B"}. This record does not assert independent reproduction. ### SafeKey Reasoning Harmful Ratio: 6. Result ID: 69a7eafb6dbb7075b94cb2da. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Qwen-7B"}. This record does not assert independent reproduction. ### STAR Reasoning Harmful Ratio: 8. Result ID: 69a7eafb6dbb7075b94cb2d2. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Llama-8B"}. This record does not assert independent reproduction. ### SafeKey Reasoning Harmful Ratio: 8.7. Result ID: 69a7eafb6dbb7075b94cb2d3. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Llama-8B"}. This record does not assert independent reproduction. ### IPO Reasoning Harmful Ratio: 11. Result ID: 69a7eafb6dbb7075b94cb2dc. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Qwen-7B"}. This record does not assert independent reproduction. ### STAR Reasoning Harmful Ratio: 11.4. Result ID: 69a7eafb6dbb7075b94cb2d9. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Qwen-7B"}. This record does not assert independent reproduction. ### RealSafe Reasoning Harmful Ratio: 18. Result ID: 69a7eafb6dbb7075b94cb2d8. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Qwen-7B"}. This record does not assert independent reproduction. ### RealSafe Reasoning Harmful Ratio: 20.7. Result ID: 69a7eafb6dbb7075b94cb2d1. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Llama-8B"}. This record does not assert independent reproduction. ### Base Reasoning Harmful Ratio: 24.3. Result ID: 69a7eafb6dbb7075b94cb2dd. [Reporting paper]() Recorded evaluation configuration: {"Backbone":"Qwen3-8B"}. This record does not assert independent reproduction. ### SafeChain Reasoning Harmful Ratio: 56.1. Result ID: 69a7eafb6dbb7075b94cb2d0. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Llama-8B"}. This record does not assert independent reproduction. ### SafeChain Reasoning Harmful Ratio: 56.1. Result ID: 69a7eafb6dbb7075b94cb2d7. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Qwen-7B"}. This record does not assert independent reproduction. ### Base Reasoning Harmful Ratio: 69. Result ID: 69a7eafb6dbb7075b94cb2cf. [Reporting paper]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Llama-8B"}. This record does not assert independent reproduction. ### Base Reasoning Harmful Ratio: 72.4. Result ID: 69a7eafb6dbb7075b94cb2d6. [Reporting paper]() Recorded evaluation configuration: {"Backbone":"DeepSeek-R1-Distill-Qwen-7B"}. This record does not assert independent reproduction. Page 1 of 1. This is an index page, not the full collection.