Google says its AI is immune to prompt injection

Google says Gemini 4 Carbon blocked every attack on the Gray Swan prompt injection benchmark, with a 0% attack success rate. But a model passing a benchmark doesn’t make every AI app built on it secure. EchoLeak (CVE-2025-32711), found by Aim Security, was the first known zero-click prompt injection exploit in a production AI system. One email, phrased like a normal message, got past Microsoft 365 Copilot’s injection classifier and made it leak data through an auto-loaded image. Microsoft patched it in 2025, with no evidence of real-world use. If you’re building AI agents: treat external content as untrusted, limit what agents can access, and validate outputs before anything leaves your system. #PromptInjection #AISecurity #Gemini #AIAgents #Cybersecurity