This research reveals how Large Language Models used for time series forecasting are vulnerable to targeted adversarial attacks, potentially compromising their reliability in critical applications.

• Black-box attack framework demonstrates how malicious actors can manipulate LLM forecasting outputs
• Minimal data perturbations can cause significant prediction errors while remaining undetected
• Security gaps identified across various time series forecasting models
• Defense mechanisms are urgently needed before widespread deployment in sensitive domains

These findings highlight a critical security concern for organizations deploying LLM-based forecasting in finance, healthcare, and infrastructure monitoring, where prediction manipulation could have severe consequences.

Adversarial Vulnerabilities in Large Language Models for Time Series Forecasting