FoodGuardBench: leaderboard
Jailbreak robustness on food safety from FoodGuardBench (2026): 2,339 harmful requests grounded in FDA food-code principles (contamination, temperature control, hygiene, storage, preparation, allergens, pest control, water safety) rewritten by the AutoDAN and PAP jailbreak methods. The score is the attack success rate, the percent of jailbreak prompts answered without a refusal phrase; lower is better.
Metric: Attack success rate (%) of the AutoDAN and PAP jailbreak versions (4,464 prompts; AutoDAN prompts optimized on Llama 3.1 8B and transferred) of the 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
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 34.59 | #470 | |
| 2 | GPT-4.1 | 50.54 | #505 | |
| 3 | GPT-4o | 54.84 | #751 | |
| 4 | Llama 3.3 70B Instruct | 55.49 | #1385 | |
| 5 | Qwen 3 8B | Non-reasoning | 56.38 | #1873 |
| 6 | GLM-4 32B | 60.84 | #1325 | |
| 7 | Qwen 3 32B | Non-reasoning | 62.76 | #972 |
| 8 | Qwen 2.5 7B | 62.99 | #2929 | |
| 9 | Mistral Small 4 | 67.58 | #1229 |
Interactive version: theaggregate.ai/benchmark?slug=foodguardbench · How It Works · Data refreshed daily, snapshot 2026-11-21.