FoodGuardBench - Vanilla Harmful Queries: leaderboard
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.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GLM-4 32B | 10 |
| 2 | Claude 3.7 Sonnet | 10.56 |
| 3 | GPT-4.1 | 10.85 |
| 4 | Qwen 3 8B (Non-reasoning) | 13.16 |
| 5 | Qwen 3 32B (Non-reasoning) | 15.9 |
| 6 | Qwen 2.5 7B | 18.42 |
| 7 | GPT-4o | 19.66 |
| 8 | Mistral Small 4 | 31.81 |
| 9 | Llama 3.3 70B Instruct | 32.65 |
Interactive version: theaggregate.ai/benchmark?slug=foodguardbench-vanilla-harmful-queries · How It Works · Data refreshed daily, snapshot 2026-10-07.