What matters in AI.

Subscribe

A food safety test shows a mean attack success rate of 25.51% in 12 LLMs

A test with one jailbreak method can underestimate the risk, as the best methods reached 60.59% on average.

Claimed, not confirmed

Researchers built FoodGuardBench, a test with 2,124 malicious food safety questions and 16,992 variants from eight jailbreak methods. In twelve LLMs, the mean attack success rate was 25.51%. General guardrails found only 55.80% of the attacks. The researchers also made FoodGuard-4B, a guardrail for food safety. Its result on new data was 97.64%.

Food safety attacks on 12 LLMs Two panels on a 0 to 100% scale. Attacks: overall mean 25.51%, best attack per model average 60.59%. Guardrails: general guardrails 55.80%, FoodGuard-4B 97.64%. Food safety attacks on 12 LLMs ATTACKS 0 100% Overall mean 25.51% Testing one attack misses this gap Best attack per model, average 60.59% GUARDRAILS 0 100% General guardrails 55.80% FoodGuard-4B 97.64% 2,124 questions, 16,992 variants, 8 jailbreak methods
Across twelve LLMs, the best attack per model averaged 60.59% success, well above the 25.51% overall mean. FoodGuard-4B beat general guardrails.

Sources

  1. Cooking Up Risks: Benchmarking and Reducing Food Safety Risks in Large Language Modelsarxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·15 OCT2026

Posted

Tags

Learn the terms in this story

Guides for this story

More in Research

All Research news