Google Gemini 4 Carbon Achieves Perfect Score on Jailbreak Benchmark

28 Oct 2026, 11:44 pm · 2h ago · 1 min read · OfficeChaiNew

Google's frontier AI model, Gemini 4 Carbon, has scored a perfect 0% vulnerability rate on Gray Swan AI's indirect prompt injection (IPI) jailbreak benchmark. The achievement highlights advanced safety engineering against adversarial manipulation and indirect injection attacks, where malicious prompts are concealed inside third-party data. The safety milestone accompanies strong performance across intelligence and capability benchmarks, including the Artificial Analysis Intelligence Index and Vending Bench. The result positions Google’s newest model as one of the most resilient enterprise-grade large language models available for mission-critical deployments requiring robust defense against prompt-based exploits.