Gemini 4 Carbon is showing up at the very top of several benchmarks — after topping Artificial Analysis Intelligence Index, Vending Bench, and Text Arena, it has now achieved a perfect score on the Gray Swan IPI jailbreak benchmark.
Google DeepMind’s flagship frontier model, Gemini 4 Carbon, has set a safety milestone by scoring a perfect 0% attack success rate on the Gray Swan Indirect Prompt Injection (IPI) benchmark. Tested across 15 attack attempts (k=15), Carbon completely shut out all automated jailbreak vectors, establishing itself as the most security-hardened large language model in the industry.
The Gray Swan IPI benchmark evaluates how resilient AI agents are against indirect prompt injection, one of the most dangerous threat vectors facing real-world LLM deployments. Unlike direct prompt injection, where an attacker explicitly instructs a model to break its safety guardrails in a conversation, indirect prompt injection occurs when malicious payload instructions are hidden inside external data that the AI agent retrieves or processes—such as third-party websites, intake forms, emails, or source code repositories. A successful attack forces the model to perform unintended harmful actions, such as leaking private user data or executing unauthorized tool calls, while tricking the model into concealing the compromise from the user.
Achieving a 0% attack success rate on this benchmark indicates that Gemini 4 Carbon maintained full structural isolation between untrusted context inputs and its system-level execution logic. Across k=1, k=10, and k=15 red-teaming attempts, attackers failed to exploit the model even once. For comparison, Gemini 4 Argon registered a 0.7% attack success rate, while top competitor Claude Opus 5.5 logged a 1.0% failure rate.
Other mainstream frontier models showed substantially higher vulnerability rates under the same 15-attempt threshold. OpenAI’s GPT-6 Astra logged an 8.5% attack success rate, Muse Spark 1.3 reached 15.9%, GLM 5.3 hit 31.5%, Grok 4.6 recorded 51.8%, and Kimi K3 saw over 52.7% of attack attempts succeed.
As enterprises deploy autonomous AI agents with access to real bank accounts, internal enterprise databases, and sensitive system tools, enterprise security has moved from a secondary metric to a primary core requirement. By completely neutralizing indirect prompt injections on the Gray Swan evaluation suite, Gemini 4 Carbon provides businesses with a model that pairs top-tier general intelligence with enterprise-grade defensive reliability.