benchmarks.bio

Biosecurity · Safeguards

BioSecBench-Refusal results

Safeguards benchmark measuring refusal behavior across harmful and routine biological requests.

Published score: Current V1 evaluation set (107 evaluations) · short horizon · data updated 2026-11-13 (November 13, 2026)

Compare current model-and-harness results for BioSecBench-Refusal. Safeguards benchmark measuring refusal behavior across harmful and routine biological requests. Current V1 covers 107 evaluations and 31 ranked configurations. Data updated 2026-11-13. The highest-ranked published configurations are Grok 4.7 (Pi, 62.4%); Muse Wonder 1.0 (Pi, 55.3%); Gemini 3.7 Flash (Pi, 54.8%). Safeguards is the trial-weighted harmonic mean of red-team refusal and routine-request compliance: a high score requires both. Refusing routine requests or complying with red-team requests lowers the score. Provider-level refusals count even when no model tokens are generated. This is separate from biological task accuracy and never contributes to Capabilities. Results describe the complete model and agent harness used for each run, and are not comparable across different benchmarks or evaluation versions.

Published BioSecBench-Refusal scores
RankModelHarnessProviderScoreVariant
1Grok 4.7PiSpaceXAI62.4%
2Muse Wonder 1.0PiMeta55.3%
3Gemini 3.7 FlashPiGoogle54.8%
4Gemini 4 ArgonPiGoogle48.7%
5Gemini 3.8 FlashPiGoogle47.4%
6Grok 4.6PiSpaceXAI46.9%
7Grok 4.6Grok BuildSpaceXAI45.6%
8GPT-6 AstraPiOpenAI40.8%
9GPT-6 SolCodexOpenAI37.9%
10GPT-6.1 SolPiOpenAI37.6%
11GPT-6 LunaCodexOpenAI37.5%
12Muse Spark OpenPiMeta34.9%
13GPT-6.1 TerraCodexOpenAI34.2%
14GPT-6.1 TerraPiOpenAI33.4%
15Claude Opus 5Claude CodeAnthropic31.8%
16GPT-6 SolPiOpenAI30.6%
17GPT-6 LunaPiOpenAI28.6%
18GPT-6 AstraCodexOpenAI25.5%
19GPT-6.1 SolCodexOpenAI25.0%
20Claude Opus 5PiAnthropic24.5%
21DeepSeek V4.1 FlashPiDeepSeek11.3%
22GLM 5.3PiZ.ai10.9%
23Kimi K3PiMoonshot8.3%
24Claude Fable 5.1PiAnthropic5.6%
25Claude Opus 5.5PiAnthropic3.8%
26Beam 1.0PiReflection3.4%
27Mistral Large 4.0PiMistral3.4%
28Claude Fable 5.1Claude CodeAnthropic0.0%
29Claude Opus 5.5Claude CodeAnthropic0.0%
30Nemotron 3 Ultra 550B A55BPiNVIDIA0.0%
31Nemotron 3.5 LightningPiNVIDIA0.0%

Methodology and provenance

Scores use a 0–100 percentage scale. See Paper PDF, Original publication, Repository for task construction, grading, aggregation, and uncertainty details.

LSVP: Anthropic's safety-filter-exempt evaluation variant, which produces fewer refusals on biology prompts. LSVP results are reported alongside the model's standard run and are not a replacement for it.