Google's Gemini AI Vulnerable to Content Manipulation

HiddenLayer researchers demonstrated that Google’s Gemini LLM (notably Gemini Pro/Ultra) is susceptible to prompt-injection and input-manipulation attacks that can leak system prompts, bypass content restrictions (jailbreaks), and exfiltrate sensitive information by using uncommon or nonsensical token sequences; these weaknesses enable generation of misinformation, disclosure of secret prompts/keys, and harmful instructions. The report notes Google’s ongoing red-teaming and mitigations but warns that consumers, enterprises using the Gemini API, and governments remain at risk until controls are strengthened.