A new test shows image safety checks do not find risk in some cultures
A judge with no view of culture gives a low risk score to dangerous prompts of one culture.
Researchers made CuRSE-Bench, a test set for the safety checks of text-to-image tools. It has 10,709 prompts that look safe in general but have a dangerous meaning in one culture. It also has 2,648 control prompts that have no such meaning. A judge with no view of culture gives a low risk score to the dangerous prompts.
How it works
The team made CultureCraft, a set of AI agents that write the prompts. The agents use only dangerous methods that two or more different sources show. Six judges, one for each region, gave a score to each of the prompts. One more judge had no view of culture and did the same.
The numbers
- The prompts are for six regions and ten types of risk.
- On the control prompts, the culture-blind judge gave the same score as the regional judges.
- Each regional judge finds the risk of its region, and not much more.
What is not known
The text gives no score numbers. The authors name a blind spot in tools in use, but the text does not name the tools. The authors made the prompts with their tool.
Insight: Safety guards read the words, not the meaning
On 21 October 2026, researchers showed that negation and opposite words get through the safety guards of 10 AI models. The tests used three languages. This test set adds culture. Maybe the two results show one problem: guards read the words, not the meaning. A test with one view of risk does not show it.
Sources
Posted