Researchers Highlight Google's Gemini AI Susceptibility to LLM Threats

HiddenLayer and other researchers identified several vulnerabilities in Google’s Gemini LLM that enable prompt injection, jailbreaking, and information leakage — including techniques using synonym attacks, repeated uncommon tokens, and malicious Google Workspace documents that can override model instructions. These issues can expose system prompts, generate harmful or illegal content, and facilitate model-stealing; Google says it runs red-team exercises and has implemented mitigations, including restrictions on election-related queries.