Across 198 community-collected prompts and nine commercial APIs, psychological manipulation outperformed technical obfuscation, attack transfer across models was limited to 16.9% of transformations, and Claude 4 models showed higher failure rates than earlier Claude versions.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
PrompTrend: Continuous Community-Driven Vulnerability Discovery and Assessment for Large Language Models
Across 198 community-collected prompts and nine commercial APIs, psychological manipulation outperformed technical obfuscation, attack transfer across models was limited to 16.9% of transformations, and Claude 4 models showed higher failure rates than earlier Claude versions.