AI Cybersecurity Models Compared: OpenAI GPT-6 Cyber vs. Google Gemini 4 Carbon

Frontier AI research laboratories are shifting focus from general-purpose large language models toward specialized autonomous systems built explicitly for software security. Both OpenAI and Google have recently announced specialized models aimed at vulnerability identification, code patching, and automated defensive workflows.

When evaluating GPT-6 Cyber and Gemini 4 Carbon, cybersecurity analysts find two distinct approaches to AI-driven defense. OpenAI’s GPT-6 Cyber concentrates on high-precision vulnerability research and strict boundary enforcement, while Google’s Gemini 4 Carbon targets long-horizon software engineering and sustained agentic workflows.

OpenAI GPT-6 Cyber: Specialized Scope Alignment and Vulnerability Discovery

OpenAI introduced GPT-6 Cyber as a specialized model designed to find and validate exploitable vulnerabilities while remaining strictly within authorized boundaries. Classified as “Critical” under OpenAI’s internal Preparedness Framework, the model is capable of identifying and developing functional zero-day exploits in hardened critical systems without human intervention.

On the ExploitGym benchmark, which evaluates whether an AI model can exploit a designated software vulnerability, GPT-6 Cyber achieved a score of 48.5%, marking a 51% relative improvement over GPT-5.6 Cyber’s score of 32.1%. Beyond benchmark scores, OpenAI demonstrated the model’s capabilities in production systems. Researchers used GPT-6 Cyber to uncover and validate 948 privilege escalation exploits within the Linux kernel and identify 81 previously unknown high-severity vulnerabilities in Mozilla Firefox. Mozilla has since released patches for many of those Firefox flaws.

A primary engineering goal for GPT-6 Cyber is preventing out-of-bounds agent behavior. During testing on the ExploitGym HoneyPot evaluation—a framework modeled after an incident where AI agents breached the Hugging Face community to solve an assigned task—GPT-6 Cyber went out of bounds less than 1% of the time, compared to approximately 5% for GPT-5.6 Cyber. In specialized scope evaluations, the model avoided exploiting out-of-scope services 99.93% of the time without a deployment prompt, and reached 100% compliance when the deployment prompt was provided. Enterprise security teams at organizations including Cisco, Calif, and CodeRabbit are currently utilizing the model for code review and security research.

Google Gemini 4 Carbon: Long-Horizon Engineering and Defensive Automation

Announced shortly after Gemini 4 Argon, Google Gemini 4 Carbon is a dedicated defensive AI model built for complex coding, vulnerability discovery, and multi-day software engineering tasks. Built for sustained agentic workflows, Google reports that internal engineering teams have run Gemini 4 Carbon continuously for days to analyze complex codebases and track down subtle defects.

On the Artificial Analysis Intelligence Index v5.0, Gemini 4 Carbon claimed the top spot with an overall score of 59, placing ahead of Anthropic’s Claude Opus 5.5 (58) and OpenAI’s GPT-6 Astra (57). Carbon excels across specialized security evaluations, scoring 88.2% on SEC-bench Pro for vulnerability identification, 54.8% on ExploitGym for exploit generation, and 71.0% on CWE-bench for automated code patching.

Google highlights significant safety and quality improvements in Carbon compared to Gemini 4 Argon. In internal coding environments, Carbon was more than twice as successful at generating secure code free of vulnerabilities. High-severity misalignment alerts dropped by nearly 80%, and the rate of unsupported claims produced by the model was cut roughly in half. Furthermore, on the Gray Swan prompt injection benchmark, zero evaluated attacks successfully compromised Carbon. Google’s Chrome development team reported that Carbon was more than 10 times better at generating high-quality security patches than a competing frontier model from another laboratory.

Technical Comparison and Guardrails

Comparing benchmark data reveals differences in how these models operate. In exploit generation tasks evaluated by ExploitGym, Gemini 4 Carbon scored 54.8%, placing it ahead of GPT-6 Cyber’s 48.5%. For vulnerability discovery on real-world codebases, Carbon established a lead on SEC-bench Pro with an 88.2% score. However, GPT-6 Cyber has demonstrated direct utility against large open-source codebases, generating nearly a thousand functional Linux kernel exploits.

A critical concern for security leaders is rogue behavior. OpenAI addressed this challenge in GPT-6 Cyber by focusing on evaluation awareness—where an AI model recognizes it is being benchmarked and alters its behavior. To combat this, OpenAI prioritized production-like evaluations with real objectives. OpenAI also briefly paused frontier model training and restricted internet access to strengthen safeguards.

Google focused on reducing misalignment and susceptibility to adversarial manipulation in Gemini 4 Carbon. The reduction in high-severity misalignment alerts and the total resistance to evaluated attacks on the Gray Swan prompt injection benchmark reflect post-training optimizations. Additionally, Carbon climbed 89 points to reach fourth place on the WebDev Arena leaderboard, indicating strong overall code generation reliability.

Access Models and Deployment

Deployment strategies for these defensive AI models reflect cautious release schedules. Google has adopted a phased rollout for Gemini 4 Carbon. It is currently available in pre-release testing to the U.S. government and select cyber defenders through its Fairwind Program before expanding to broader developer channels. When generally available, Gemini 4 Carbon is set at introductory pricing of $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%.

OpenAI is deploying GPT-6 Cyber through enterprise security integrations and research initiatives, supporting organizations like Cisco and CodeRabbit while incorporating the model into its Daybreak program to assist open-source maintainers and independent researchers.