# Safety & robustness benchmarks > Source: Benchize (https://benchize.com) — https://benchize.com/benchmarks/category/safety. When citing or reusing this content, credit "Benchize" and link to https://benchize.com/benchmarks/category/safety. Explore Safety & robustness LLM benchmarks and AI model results. Compare source-reported scores with original metrics, evaluation settings and evidence on Benchize. Canonical URL: https://benchize.com/benchmarks/category/safety Last updated: 2026-10-21T18:52:20.672Z Reproduction and quotation must name Benchize, link to the canonical HTML URL, and retain individual source citations. Explore safety, security, truthfulness, and robustness under adversarial inputs. Read each metric's direction carefully: higher refusal or attack-success rates can have very different meanings. Published Oct 21, 2026 ## [CyberBench v1.1 / overall](https://benchize.com/benchmarks/benchmark%3Acyber%3Aoverall.md) Updated: 10/21/2026 Models tested: 28 Can autonomous agents craft PoC inputs that trigger OSS-Fuzz vulnerabilities—and produce source patches that fix them? Top models · Vals AI Latest result per model · 1.1 - 1[Muse Wonder](https://benchize.com/models/muse-wonder-max.md)Max - 2[GPT-6 Sol](https://benchize.com/models/gpt-6-sol-max.md)Max - 3[Gemini 4 Argon](https://benchize.com/models/gemini-4-argon-high.md)High View details → ## [CyberBench v1.1 / patch](https://benchize.com/benchmarks/benchmark%3Acyber%3Apatch.md) Updated: 10/21/2026 Models tested: 28 Can autonomous agents craft PoC inputs that trigger OSS-Fuzz vulnerabilities—and produce source patches that fix them? Top models · Vals AI Latest result per model · 1.1 - 1[GPT-5.6 Sol](https://benchize.com/models/gpt-5-6-sol-max.md)Max - 2[Claude Haiku 5.5](https://benchize.com/models/claude-haiku-5-5-max.md)Max - 3[GPT-6 Sol](https://benchize.com/models/gpt-6-sol-max.md)Max View details → ## [CyberBench v1.1 / poc](https://benchize.com/benchmarks/benchmark%3Acyber%3Apoc.md) Updated: 10/21/2026 Models tested: 28 Can autonomous agents craft PoC inputs that trigger OSS-Fuzz vulnerabilities—and produce source patches that fix them? Top models · Vals AI Latest result per model · 1.1 - 1[Muse Wonder](https://benchize.com/models/muse-wonder-max.md)Max - 2[Gemini 4 Argon](https://benchize.com/models/gemini-4-argon-high.md)High - 3[GPT-6 Sol](https://benchize.com/models/gpt-6-sol-max.md)Max View details → ## [DistressBench](https://benchize.com/benchmarks/scale%3Adistressbench.md) Updated: 10/21/2026 Models tested: 27 Conversational models have become a default place people turn during anxiety, grief, and acute distress. A growing share of that use is emotional, and within it users disclose… Top models · scale Latest result per model · Version not reported - 1[Muse Spark 1.2](https://benchize.com/models/muse-spark-1-2.md) - 2[Muse Spark 1.3](https://benchize.com/models/muse-spark-1-3.md) - 3[Muse Glimmer 30B](https://benchize.com/models/muse-glimmer-30b.md) View details → ## [Fortress](https://benchize.com/benchmarks/scale%3Afortress.md) Updated: 10/21/2026 Models tested: 31 The rapid advancement of large language models (LLMs) introduces powerful dual-use capabilities that could both threaten and bolster national security and public safety (NSPS).… Top models · scale Latest result per model · Version not reported - 1[DeepSeek R1](https://benchize.com/models/deepseek-r1.md) - 2[Qwen3 235B A22B](https://benchize.com/models/qwen3-235b-a22b.md) - 3[DeepSeek V3.1](https://benchize.com/models/deepseek-v3-1.md) View details → ## [MASK](https://benchize.com/benchmarks/scale%3Amask.md) Updated: 10/21/2026 Models tested: 25 As AI systems become more powerful, autonomous, and agentic, users need confidence that these systems won't deceive them. A model that regularly produces statements contradicting… Top models · scale Latest result per model · Version not reported - 1[Claude Opus 4.6](https://benchize.com/models/claude-opus-4-6-off.md)Non-thinking - 2[Claude Sonnet 4](https://benchize.com/models/claude-sonnet-4-thinking.md)Thinking - 3[gpt-oss-120b](https://benchize.com/models/gpt-oss-120b.md) View details → ## [Spiral-Bench v1.2](https://benchize.com/benchmarks/eq-bench%3Aspiral-bench-v1.2.md) Updated: 10/21/2026 Models tested: 16 Safety and social handling of simulated twenty-turn conversations with users vulnerable to delusional thinking. An ensemble of judges grades protective and risky behaviors; v1.2… Top models · eq bench Latest result per model · spiral-bench-v1.2 - 1[Claude Sonnet 4.5](https://benchize.com/models/claude-sonnet-4-5.md) - 2[GPT-5.2](https://benchize.com/models/gpt-5-2.md) - 3[O3](https://benchize.com/models/o3.md) View details → ## [Spiral-Bench v1.0](https://benchize.com/benchmarks/eq-bench%3Aspiral-bench-v1.0.md) Updated: 10/21/2026 Models tested: 11 The original Spiral-Bench protocol for simulated conversations with users vulnerable to delusional thinking. A single judge grades protective and risky behaviors; these legacy… Top models · eq bench Latest result per model · spiral-bench-v1.0 - 1[O3](https://benchize.com/models/o3.md) - 2[gpt-oss-120b](https://benchize.com/models/gpt-oss-120b.md) - 3[Claude Sonnet 4.5](https://benchize.com/models/claude-sonnet-4-5.md) View details → ## [PropensityBench](https://benchize.com/benchmarks/scale%3Apropensitybench.md) Updated: 10/21/2026 Models tested: 5 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… Top models · scale Latest result per model · Version not reported - 1[Gemini 2.5 Pro](https://benchize.com/models/gemini-2-5-pro.md) - 2[Gemini 2.0 Flash](https://benchize.com/models/gemini-2-0-flash.md) - 3[Qwen3 8B](https://benchize.com/models/qwen3-8b.md) View details → ## [Adversarial Robustness](https://benchize.com/benchmarks/scale%3Aadversarial_robustness.md) Updated: 10/21/2026 Models tested: 4 The Scale Adversarial Robustness Evaluation runs on 1,000 human-written prompts that are designed to trigger harmful responses from large language models. These prompts cover… Top models · scale Latest result per model · Version not reported - 1[Mistral Large](https://benchize.com/models/mistral-large.md) - 2[GPT-4 Turbo Preview](https://benchize.com/models/gpt-4-turbo-preview.md) - 3[Gemini 1.5 Flash](https://benchize.com/models/gemini-1-5-flash.md) View details → ## Methodology and provenance Benchize collects and structures public benchmark reports. Scores, costs, timings, ranks and uncertainty remain attributed to their original sources. Different versions, configurations and measurement units are separate results. Missing fields stay unavailable. --- Cite as: Safety & robustness benchmarks — Benchize, https://benchize.com/benchmarks/category/safety (retrieved 2026-10-21)