OpenAI launches GPT-6 Cyber, uncovering 81 high-severity flaws in Firefox and hundreds of Linux kernel bugs
OpenAI has officially launched GPT-6 Cyber, a specialized variant of its frontier GPT-6 architecture trained specifically for offensive and defensive cybersecurity research. The model, built to autonomously discover software vulnerabilities, validate exploits, and generate candidate patches, immediately demonstrated its capabilities by uncovering 81 novel, high-severity zero-day flaws in Mozilla Firefox. For context, this is one of the world's most heavily audited open-source codebases.
Due to the system's unprecedented ability to execute complex, multi-step cyber operations, OpenAI categorized GPT-6 Cyber at the "Critical" risk tier under its internal Preparedness Framework and is gating access strictly through its vetted Daybreak Red partner program.
Real-World Audit Campaigns: Firefox and the Linux Kernel
OpenAI deployed GPT-6 Cyber across major open-source infrastructure projects to stress-test its autonomous vulnerability discovery and end-to-end proof-of-concept (PoC) generation.
Mozilla Firefox Audit
The model identified 81 previously unknown vulnerabilities classified by Mozilla as high-severity, while demonstrating that an additional 30 known issues carried higher operational risks than previously recognized. OpenAI provided Mozilla with reproducible test cases and candidate code patches, allowing Firefox engineers to ship fixes to millions of users prior to the public announcement.
Linux Kernel Exploitation
In systematic research targeting the Linux kernel, GPT-6 Cyber reproduced 1,724 vulnerabilities and successfully developed working exploits for 948 of them. That’s more than five times the output of predecessor models using identical research harnesses. OpenAI submitted 304 critical findings and patches directly to kernel maintainers.
Web Penetration Testing
On OpenAI's black-box web penetration benchmark, which requires agents to audit applications without source code visibility, GPT-6 Cyber achieved a 40.3% vulnerability recall rate, outpacing base frontier models like GPT-6 Astra (32.0%).
Alignment and Anti-Spillover Safeguards
Given the model's offensive efficiency, OpenAI focused heavily on boundary compliance, ensuring the agent refrains from attacking out-of-scope targets or escalating actions during authorized security engagements.
Exploitation Guardrails: Evaluated on the ExploitGym benchmark - a test built after prior AI agents accidentally breached out-of-scope infrastructure during sandbox tests - GPT-6 Cyber achieved an exploit success rate of 48.5% while adhering to target boundaries.
Zero Out-of-Bounds Drift: With deployment safety instructions enabled in its chain-of-thought system, the model maintained a 0.0% out-of-bounds attack rate, successfully halting actions when reaching unauthorized network margins or third-party APIs.
Tiered Deployment and the Offense-Defense Race
OpenAI is restricting initial access to approved enterprise security teams, critical infrastructure operators, and defensive research entities via its Daybreak Red API and Codex environments. API requests are locked to verified IP networks by default to prevent token hijacking or rogue use.
The release accentuates a central debate across the cybersecurity ecosystem: Can gating autonomous vulnerability-hunting tools allow defenders to patch legacy codebases faster than adversary groups can reverse-engineer similar agentic models to launch automated zero-day attacks at scale?
