# PropensityBench > Source: Benchize (https://benchize.com) — https://benchize.com/benchmarks/scale%3Apropensitybench. When citing or reusing this content, credit "Benchize" and link to https://benchize.com/benchmarks/scale%3Apropensitybench. PropensityBench is a novel benchmark framework designed to evaluate the latent safety risks in LLMs. Traditional safety evaluations primarily test a model'scapabilities, leaving a critical blind spot regarding a model'spropensity, or what itwould doif empowered with dangerous tools alongside safe ones. Models may strategically conceal risky capabilities or harbor latent inclinations toward misuse that static audits fail to capture. PropensityBench addresses this by shifting the focus from "can-do" to "would-do" assessments. It measures an LLM's tendency to pursue harmful actions when given the choice in a simulated, high-stakes environment. The framework spans four critical, high-risk domains:biosecurity, chemical security, cybersecurity, and self-proliferation. By applying realistic operational pressures, the benchmark reveals a model's underlying inclinations, providing a more dynamic and realistic picture of its true safety alignment. Canonical URL: https://benchize.com/benchmarks/scale%3Apropensitybench Last updated: Wed, 21 Oct 2026 18:52:20 GMT Reproduction and quotation must name Benchize, link to the canonical HTML URL, and retain individual source citations. PropensityBench is a novel benchmark framework designed to evaluate the latent safety risks in LLMs. Traditional safety evaluations primarily test a model'scapabilities, leaving a critical blind spot regarding a model'spropensity, or what itwould doif empowered with dangerous tools alongside safe ones. Models may strategically conceal risky capabilities or harbor latent inclinations toward misuse that static audits fail to capture. PropensityBench addresses this by shifting the focus from "can-do" to "would-do" assessments. It measures an LLM's tendency to pursue harmful actions when given the choice in a simulated, high-stakes environment. The framework spans four critical, high-risk domains:biosecurity, chemical security, cybersecurity, and self-proliferation. By applying realistic operational pressures, the benchmark reveals a model's underlying inclinations, providing a more dynamic and realistic picture of its true safety alignment. Safety & robustness · Models tested: 5 · Score publishers: [scale](https://labs.scale.com/leaderboard/propensitybench) ## Model results | # | Model | Score | ± | | | --- | --- | --- | --- | --- | | 1 | [Gemini 2.5 Pro](https://benchize.com/models/gemini-2-5-pro.md) | 79source score | CI 78.4–79.6 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 79 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration gemini-2.5-pro sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 78.4–79.6 contaminationWarning Not reported rank 14 reportedConfidenceHalfWidth 0.6 ### Source [Original source](https://labs.scale.com/leaderboard/propensitybench) Source scale Original model label gemini-2.5-pro Observed 2026-10-21T16:35:35.949Z Evidence SHA-256 `e65bbee55f6a735247cf59c5746c7850ce3413b5c0686e4c5b71a8cd641ec37f` Source locator `/rows/13` | | | | | | 2 | [Gemini 2.0 Flash](https://benchize.com/models/gemini-2-0-flash.md) | 77.8source score | CI 77–78.6 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 77.8 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration gemini-2.0-flash sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 77–78.6 contaminationWarning Not reported rank 13 reportedConfidenceHalfWidth 0.8 ### Source [Original source](https://labs.scale.com/leaderboard/propensitybench) Source scale Original model label gemini-2.0-flash Observed 2026-10-21T16:35:35.949Z Evidence SHA-256 `e65bbee55f6a735247cf59c5746c7850ce3413b5c0686e4c5b71a8cd641ec37f` Source locator `/rows/12` | | | | | | 3 | [Qwen3 8B](https://benchize.com/models/qwen3-8b.md) | 75.2source score | CI 75.1–75.3 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 75.2 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration Qwen3-8B sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 75.1–75.3 contaminationWarning Not reported rank 12 reportedConfidenceHalfWidth 0.1 ### Source [Original source](https://labs.scale.com/leaderboard/propensitybench) Source scale Original model label Qwen3-8B Observed 2026-10-21T16:35:35.949Z Evidence SHA-256 `e65bbee55f6a735247cf59c5746c7850ce3413b5c0686e4c5b71a8cd641ec37f` Source locator `/rows/11` | | | | | | 4 | [Gemini 2.5 Flash](https://benchize.com/models/gemini-2-5-flash.md) | 68source score | CI 67.8–68.2 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 68 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration gemini-2.5-flash sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 67.8–68.2 contaminationWarning Not reported rank 11 reportedConfidenceHalfWidth 0.2 ### Source [Original source](https://labs.scale.com/leaderboard/propensitybench) Source scale Original model label gemini-2.5-flash Observed 2026-10-21T16:35:35.949Z Evidence SHA-256 `e65bbee55f6a735247cf59c5746c7850ce3413b5c0686e4c5b71a8cd641ec37f` Source locator `/rows/10` | | | | | | 5 | [GPT-4o (2024-11-20)](https://benchize.com/models/gpt-4o-2024-11-20.md) | 46.1source score | CI 45.7–46.5 | | | ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 46.1 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort unknown Model configuration gpt-4o-2024-11-20 sourceModelVersion Not reported thinkingEnabled Not reported ### Reported statistics Uncertainty Confidence interval: 45.7–46.5 contaminationWarning Not reported rank 7 reportedConfidenceHalfWidth 0.4 ### Source [Original source](https://labs.scale.com/leaderboard/propensitybench) Source scale Original model label gpt-4o-2024-11-20 Observed 2026-10-21T16:35:35.949Z Evidence SHA-256 `e65bbee55f6a735247cf59c5746c7850ce3413b5c0686e4c5b71a8cd641ec37f` Source locator `/rows/6` | | | | | --- Cite as: PropensityBench — Benchize, https://benchize.com/benchmarks/scale%3Apropensitybench (retrieved 2026-10-21)