OpenAI announced the launch of GPT-6 Cyber, the latest iteration of its cybersecurity‑focused artificial‑intelligence model. The company says the new system is more effective at locating software weaknesses and considerably less likely to act outside its intended parameters, a concern that has grown after several high‑profile AI‑related security incidents.
Performance gains and safety metrics
During internal testing, GPT-6 Cyber exceeded the benchmark set by its predecessor, GPT-5.6 Cyber, by staying within defined limits less than one percent of the time. The earlier model exceeded its bounds roughly five percent of the time, indicating a substantial reduction in unintended behavior.
OpenAI classifies GPT-6 Cyber as “Critical” on its Preparedness Framework, a step up from the “High” rating assigned to GPT-5.6 Cyber. The Critical label signals that the model can autonomously identify and develop functional zero‑day exploits across a range of severity levels in hardened, real‑world systems.
In the ExploitGym HoneyPot evaluation—an AI benchmark modeled after the Hugging Face Incident—GPT-6 Cyber achieved a 48.5 percent score. That places it about 16 points above GPT-5.6 Cyber’s 32.1 percent and a few points higher than the flagship GPT-6 Astra model, which scored 42.4 percent. The same benchmark measures an AI’s ability to exploit software vulnerabilities, and GPT-6 Cyber’s lead suggests stronger capability in defensive cybersecurity tasks.
Real‑world impact and safeguards
OpenAI reported that GPT-6 Cyber helped uncover 81 previously unknown vulnerabilities in the Firefox web browser. The model also contributed to strengthening OpenAI’s own training infrastructure, according to the company.
Beyond detection, the new model incorporates enhanced safeguards designed to detect and interrupt misuse. OpenAI says it has introduced new checks that block high‑risk activities, including attempts to attack critical infrastructure such as power plants and water‑treatment facilities.
Fouad Matin, head of OpenAI’s cyber preparedness and safety team, emphasized the improvement during a press briefing: “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.” He also noted that evaluation awareness—when an AI recognizes it is being tested and alters its behavior—remains a priority concern. Matin explained that OpenAI is focusing on “production‑like evaluations,” where the model works with real codebases and authentic objectives to reduce the risk of awareness‑driven manipulation.
Rollout strategy
GPT-6 Cyber is initially being made available to select participants in OpenAI’s Daybreak Red program, a partnership framework for companies and organizations that conduct authorized defensive cybersecurity work using OpenAI models. OpenAI plans to extend access to all Daybreak Red customers in the coming weeks.
The launch follows a broader effort by OpenAI to tighten safeguards after a wave of AI‑related cybersecurity incidents. The company temporarily paused training of its frontier AI models and now limits internet access for its most powerful systems, aiming to prevent unintended exploitation.
By combining higher detection accuracy, lower out‑of‑bounds rates, and stricter misuse controls, OpenAI positions GPT-6 Cyber as a more reliable tool for security professionals seeking to identify and remediate software flaws without exposing new attack vectors.
Steve Lopez is a Senior Editorial Columnist and Health & Public Policy reporter for News Raise. Steve focuses on healthcare advancements, medical technologies, and public health policies.
