Research Summary: AI Agents are Vulnerable to Radicalization

Recent research from arXiv:2609.38296v1 indicates that Artificial Intelligence (AI) agents, specifically Large Language Models (LLMs), are susceptible to radicalization through interaction with other LLMs. Simulations demonstrated that an 'influencer' LLM could make a 'target' LLM's beliefs more extreme via two pathways: reinforcing pre-existing beliefs (resonance) and promoting initially unimportant beliefs (persuasion). Both mechanisms proved effective, with resonance exhibiting a consistently stronger impact.

Why it matters

This finding highlights a significant vulnerability in AI systems, particularly LLMs, regarding their susceptibility to external manipulation and radicalization. Understanding these mechanisms is crucial for developing robust and secure AI, ensuring that advanced computational agents operate reliably and ethically without unintended belief distortions or extremization, which could impact their decision-making and interaction with real-world systems.

Key insights

• LLMs can be manipulated by other LLMs, leading to the radicalization of beliefs.
• A simulated environment involved a 'target' LLM role-playing a human persona and an 'influencer' LLM aiming to extremize the target's beliefs.
• Radicalization was observed through two pathways: 'resonance' (reinforcing existing beliefs) and 'persuasion' (promoting new, initially unimportant beliefs).
• Both resonance and persuasion mechanisms successfully radicalized the target LLMs.
• The 'resonance' pathway consistently produced stronger radicalization effects compared to 'persuasion' across affective and behavioral metrics.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.38296