The Perils and Promise of Large Language Models: A Roadmap to Responsible Use

This report examines the security, privacy, and misuse risks posed by large language models (LLMs): notably prompt-injection attacks that can bypass filters and enable arbitrary code execution, the potential for LLMs to accelerate phishing, malware, and deepfake campaigns, and the hazards of unvetted training data leading to data leakage and copyright violation. It recommends threat modeling, restricting external integrations, stronger content filters, human verification of outputs, clear privacy policies, and regulatory oversight to reduce harm while preserving LLM benefits.