How Far Will They Go? Red-Teaming Online Influence with Large Language Models

In three linesRed-teaming study of 30+ open-source LLMs (10 families, 5 countries) measuring capacity to generate biased political content via jailbreaks. Findings: systematic asymmetries (left-leaning bias), Overton Window contraction with model size, substantial regional differences, variable jailbreak potency across model families.

Summary generated by Claude — human-verified