The British Institute Critiques the Consistency of AI Security Tests

British researchers present a method for identifying language models that are more cautious during testing than in everyday use. They warn that systematically blocking certain queries can artificially inflate security scores, to the detriment of utility. Their study, conducted at the AI Security Institute using psychometric methods, questions the consistency of popular security benchmarks.

A Method to Detect More Cautious Models in Testing

The researchers propose a method to identify language models that exhibit more cautious behavior during evaluations than in regular use. They indicate that systematically blocking certain queries can lead to an artificial increase in the security score, even if it reduces the model's utility in daily applications. Their approach aims to highlight a behavioral gap between the testing phase and normal usage.

Security Benchmarks Called into Question

Researchers at the AI Security Institute in the UK have used psychometric methods to analyze the security benchmarks applied to language models. According to them, these benchmarks, while widely used, do not assess a consistent trait. Their analysis highlights significant weaknesses in how the security of language models is measured.