Reproduce “Persona Jailbreaking in Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jivnesh Sandhan, Fei Cheng, Tushar Sandhan, Yugo Murawaki
Researchers from Kyoto University and IIT Kanpur uncovered a vulnerability termed "persona editing" in Large Language Models, demonstrating how adversarial conversational history can subtly and systematically alter an LLM's Big Five personality traits using their PHISH framework, often without impairing core reasoning abilities.
Key insights
- LLM personas are contextually emergent at inference and highly susceptible to implicit adversarial steering via conversational history, even without explicit refusal bypass.
- Successful persona hijacking is primarily driven by reverse-polarity answers and trait-specific question framing, rather than general in-context learning or reasoning elements.
- LLMs exhibit stronger inter-trait correlations in their Big Five profiles compared to human psychological patterns, suggesting an overly coupled encoding of personality dimensions that results in spillover effects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.