The authors constructed a benchmark of 260 prompts covering a range of harmful requests, each annotated for harmfulness potential on a 0–5 scale.
Polarity blindness causes LLM safety guardrails to miss harmful intent after negation
Across 10 LLMs and three languages, negation and antonym replacement produce consistently higher harm scores than synonym changes, with a mean harm score of 4.12 for the most vulnerable model.
Academic
Diego Miguel Lozano · Erik Derner · Juan Antonio Pérez-Ortiz · Nuria Oliver
ELLIS Alicante · Czech Technical University in Prague · Universidad de Alicante
Research Digest··3 min read
The authors systematically probe LLMs with prompts modified via negation, antonym substitution, and synonym substitution (as a control), both with and without an adversarial textual wrapper.
What it released
Code
§