OpenAI has introduced GPT-6 Cyber, a new cybersecurity model classified as 'Critical,' which demonstrates superior performance on cyber-related benchmarks compared to the GPT-6 Astra.
OpenAI claims that GPT-6 Cyber is significantly better at identifying vulnerabilities while minimizing the risk of the model acting maliciously against other platforms. During testing, the model remained within safe boundaries less than 1 percent of the time, a significant improvement over the previous GPT-5.6 Cyber, which had a rate of approximately 5 percent.
The evaluation utilized a framework named ExploitGym HoneyPot, inspired by an incident where OpenAI's AI agents breached the Hugging Face community. Fouad Matin, the head of OpenAI’s cyber preparedness and safety team, highlighted these advancements during a press briefing, stating, "With the 6 Cyber model, we were able to make huge improvements on this evaluation. It’s over 99 percent of the time doing what you ask."
Despite these improvements, the company remains cautious about "evaluation awareness," where an AI system alters its behavior after recognizing it is being tested. Matin noted that this issue often appears in the model's chain of thought, leading OpenAI to prioritize production-like evaluations that involve real codebases and objectives to mitigate this risk.
In response to a series of recent cybersecurity incidents involving AI, OpenAI has accelerated efforts to enhance safeguards and restrict access to the internet for its most powerful frontier models.
GPT-6 Cyber is categorized as 'Critical' under the Preparedness Framework, signifying its ability to identify and develop functional zero-day exploits in hardened systems without human intervention. This classification supersedes the previous 'High' rating given to GPT-5.6 Cyber.
The model performs better than GPT-6 Astra on most cyber tasks, achieving a score of 48.5 percent in the ExploitGym benchmark, compared to 42.4 percent for Astra and 32.1 percent for GPT-5.6 Cyber. Notably, GPT-6 Cyber discovered 81 previously unknown vulnerabilities in Firefox and contributed to securing OpenAI’s own training infrastructure.
Alongside its enhanced capabilities, GPT-6 Cyber includes stricter safety measures designed to detect and prevent misuse, particularly blocking attacks on critical infrastructure such as power plants and water treatment facilities.
Availability is currently limited to select members of the Daybreak Red program, which is reserved for organizations conducting authorized defensive cybersecurity work. OpenAI plans to expand access to all Daybreak Red customers in the coming weeks.
The Editorial Staff at AIChief is a team of professional content writers with extensive experience in AI and marketing. Founded in 2025, AIChief has quickly grown into the largest free AI resource hub in the industry.