Food safety is a practical concern for AI-assisted meal planning: a PwC 2025 consumer survey cited in the paper reports that 47% of respondents are comfortable using generative AI for meal planning. Yet the safety of large language models against malicious food-related requests has been underexplored. To close this gap, the authors introduce FoodGuardBench, an FDA-grounded benchmark containing 2,124 seed malicious questions, 16,992 variants generated from eight jailbreak strategies, and 1,000 benign questions spanning five food-safety risk categories.
Evaluating twelve LLMs, the researchers found a mean overall attack success rate (ASR) of 25.51%, but the average per-model maximum across the eight attacks was 60.59%. This wide gap indicates that testing only a single attack strategy can substantially underestimate the full extent of food-safety risk. For several models, nearly half or more of the unsafe answers were rated as ready-to-use guides for harmful actions.
The paper also shows that existing general-purpose guardrails are not sufficient: they averaged 96.25% benign utility but only 55.80% detection accuracy against the eight jailbreak attacks. Using FoodGuardBench, the authors fine-tuned FoodGuard-4B, a food-safety-specific guardrail that achieves 97.64% out-of-distribution accuracy while maintaining high benign utility. This result demonstrates the benchmark's value for both identifying and mitigating food-safety risks in LLMs.