Reproduce “Exploiting Large Language Models (LLMs) through Deception Techniques and Persuasion Principles” 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
Sonali Singh, Faranak Abri, Akbar Siami Namin
An investigation into Large Language Models (LLMs) found that they can be manipulated by social engineering techniques derived from human psychological persuasion principles, prompting them to generate information for potentially malicious ends despite their robust resistance to direct harmful queries.
Key insights
- Current LLM safety mechanisms are effective against explicit malicious queries but are vulnerable to sophisticated social engineering tactics based on human persuasion principles.
- LLMs' inherent helpfulness and contextual understanding can be exploited by psychological manipulation, leading them to generate content that facilitates adversarial activities.
- There are varying levels of resilience among different LLMs, with some models (e.g., GPT4, BARD) showing greater susceptibility to deception than others (e.g., Llama 2, Claude).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.