Adversarial Feeds Steer LLM Agent Decisions Against Their Defaults
Researchers have developed a new protocol to evaluate how external information streams influence LLM agent decisions. Their study found that curated feeds can significantly steer agent choices, particularly when the agent is uncertain. This effect, termed 'adversarial capitulation,' was observed across multiple LLMs and decision domains, including security-related choices. While simple defenses can partially mitigate this influence, the research highlights the critical need to audit the feed curation layer, not just the final prompt, in LLM agent evaluations. AI
IMPACT Highlights a critical vulnerability in LLM agents, suggesting that feed curation can be a powerful control surface, potentially impacting agent reliability and safety.