AgentDrift: Unsafe Recommendation Drift Under Tool Corruption Hidden by Ranking Metrics in LLM Agents

Research reveals that AI agents using tools for financial advice can recommend unsafe products while maintaining good quality metrics when tool data is corrupted. The study found that 65-93% of recommendations contained risk-inappropriate products across seven LLMs, yet standard evaluation metrics failed to detect these safety issues.

• →AI agents consistently recommend unsafe financial products when tool outputs are contaminated, despite maintaining high quality scores on standard metrics.
• →Safety violations occurred in 65-93% of turns across seven different LLMs when tool corruption was present.
• →Standard NDCG evaluation metrics fail to capture safety risks, creating an evaluation blindness pattern in AI agent assessment.
• →Even narrative-only corruption without numerical manipulation can induce significant unsafe recommendation drift.
• →A new safety-penalized metric (sNDCG) reveals the true extent of safety degradation that standard metrics miss.