LLM Jailbreaking Risks in Smart Grid Operations

A recent study investigates vulnerabilities related to jailbreaking in large language models (LLMs) used as assistants in smart grid operations, examining three specific models under adversarial conditions. The models tested include OpenAI's GPT-4o mini, Google’s Gemini 2.0 Flash-Lite, and Anthropic's Claude 3.5 Haiku, evaluated using Baseline, BitBypass, and DeepInception techniques across nine NERC Reliability Standards scenarios. The findings revealed an overall Attack Success Rate (ASR) of 33.1%, with DeepInception reaching an ASR of 63.17%. Notably, Claude 3.5 Haiku demonstrated complete immunity with a 0% ASR. The study underscores the dangers posed by authorized users creating harmful prompts to circumvent safety measures and generate non-compliant instructions.