OpenAI Unveils GPT-6 Cyber After Model Finds 81 High-Severity Firefox Flaws
Research laboratory OpenAI has granted authorized cybersecurity teams access to GPT-6 Cyber, an advanced artificial intelligence model designed to discover software vulnerabilities. Company researchers reported using the specialized model to uncover 81 previously unknown vulnerabilities inside the Mozilla Firefox web browser. The software maintainers at Mozilla classified the discovered browser flaws as high-severity vulnerabilities and deployed remediation patches. Firefox distinguished engineer Brian Grinstead stated that the automated system uncovered critical weaknesses even after developers hardened code using earlier frontier models. The technical findings accompany a controlled distribution conducted through Daybreak Red, OpenAI's access program for authorized security evaluations.
Vulnerability Research and Linux Kernel Exploits
In addition to browser assessments, OpenAI researchers utilized GPT-6 Cyber to develop functioning proof-of-concept exploits for 948 Linux kernel vulnerabilities. Demonstrating working exploit mechanics allowed software analysts to understand precisely how attackers could potentially compromise vulnerable systems. Following open-source disclosure guidelines, company researchers reported 304 critical findings directly to Linux kernel maintainers alongside proposed remediation code. Established security enterprises including Cisco, Calif, and CodeRabbit have begun deploying the specialized system for code audits and vulnerability research. In standardized ExploitGym evaluations, GPT-6 Cyber achieved an exploit demonstration score of 48.5%, compared with 32.1% recorded by predecessor GPT-5.6 Cyber.
Safety Classification and Red Team Governance
Under OpenAI's internal Preparedness Framework, evaluation teams elevated the model's cybersecurity capability rating to Critical from its earlier High threshold. Unlike general-purpose baseline models such as GPT-6 Astra, the cyber-specialized system is configured to process exploit simulation requests from authorized defenders. Company engineers implemented stricter operational guardrails to ensure that autonomous software agents strictly observe authorized testing boundaries during engagements. In honeypot security simulations, the autonomous agent avoided interacting with unauthorized out-of-scope services across 99.93% of automated testing trials. The company is concurrently widening its companion Daybreak Blue initiative to support verified organizational defenders conducting vulnerability verification and patch validation. Mainline models such as GPT-6 Astra already handle code review, vulnerability triage and patch development, while GPT-6 Cyber is offered through Daybreak Red for deeper exploit development and authorized security testing, according to OpenAI.