FoodGuardBench - Vanilla Harmful Queries: leaderboard

Refusal of plainly worded harmful food-safety requests from FoodGuardBench (2026): 2,339 queries grounded in FDA food-code principles across eight hazard categories, asked without any jailbreak. The score is the attack success rate, the percent of requests answered without a refusal phrase; lower is better.

Metric: Attack success rate (%) of the plain, non-jailbroken harmful food-safety queries of FoodGuardBench (2,339 vanilla queries grounded in FDA food-code principles across eight hazard categories), temperature 0.7, a response counting as a successful attack when it contains none of a curated set of refusal prefixes; lower is better. Source: arxiv.org. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GLM-4 32B10#1325
2Claude 3.7 Sonnet10.56#470
3GPT-4.110.85#505
4Qwen 3 8BNon-reasoning13.16#1873
5Qwen 3 32BNon-reasoning15.9#972
6Qwen 2.5 7B18.42#2929
7GPT-4o19.66#751
8Mistral Small 431.81#1229
9Llama 3.3 70B Instruct32.65#1385

Interactive version: theaggregate.ai/benchmark?slug=foodguardbench-vanilla-harmful-queries · How It Works · Data refreshed daily, snapshot 2026-11-21.