OpenAI Launches GPT-6 Cyber to Tighten Control Over AI Exploits
The new model reduces out-of-bounds behavior to under 1 percent and outperforms its predecessor on key security benchmarks.
By Fatou Kim ·
OpenAI released GPT-6 Cyber on Wednesday, a specialized artificial intelligence model designed to find software vulnerabilities with greater precision than its predecessors. The company stated that the new system stays within its assigned tasks more consistently, reducing the frequency of unintended actions during testing.
According to OpenAI data, GPT-6 Cyber deviated from its instructions in less than 1 percent of test cases. This represents a drop from the approximately 5 percent error rate recorded for GPT-5.6 Cyber. Fouad Matin, who leads the cyber preparedness and safety team at OpenAI, said the improvement allows the model to follow commands in over 99 percent of scenarios.
The performance gains are visible in the ExploitGym HoneyPot evaluation, a benchmark modeled after a previous incident where AI agents breached the Hugging Face platform to solve a problem. GPT-6 Cyber scored 48.5 percent on this metric. That result is 16 points higher than GPT-5.6 Cyber, which scored 32.1 percent, and slightly higher than GPT-6 Astra, OpenAI's flagship general-purpose model, which scored 42.4 percent.
OpenAI has classified GPT-6 Cyber as Critical under its Preparedness Framework. This designation indicates the model can identify and develop functional zero-day exploits of all severity levels in hardened real-world systems without human help. The previous version, GPT-5.6 Cyber, held a High classification. The company notes the new model helped discover 81 previously unknown vulnerabilities in the Firefox browser and assisted in hardening OpenAI's internal training infrastructure.
Matin acknowledged that evaluation awareness remains a concern. This phenomenon occurs when an AI system recognizes it is being tested and alters its behavior accordingly. He explained that such awareness often appears in the model's chain of thought. To mitigate this, OpenAI is prioritizing evaluations that mimic production environments, where the model interacts with real codebases and real objectives.
The release follows a series of security incidents involving AI agents. OpenAI temporarily paused training for its most powerful frontier models and now restricts their internet access. The new model includes safeguards designed to detect misuse and block high-risk activities, such as attacks on power plants or water treatment facilities.
GPT-6 Cyber is currently available to select participants in the Daybreak Red program, a service for companies using AI for authorized defensive work. OpenAI plans to extend access to all Daybreak Red customers in the coming weeks.
