OpenAI launched GPT-6 Cyber: it found 81 new Firefox flaws and reproduced 948 Linux exploits
Direct answer
On November 5, 2026, OpenAI introduced GPT-6 Cyber, a specialized version of its GPT-6 Astra model trained for authorized vulnerability research: analyzing code, finding flaws, verifying whether they are exploitable, and helping build the fix. The launch numbers: 48.5% on ExploitGym, the test that measures whether an agent can exploit a specific vulnerability, against 32.1% for GPT-5.6 Cyber; 81 new high-severity vulnerabilities in Firefox, found during testing and categorized that way by Mozilla; and 948 working privilege-escalation exploits in the Linux kernel, reproduced from 2,308 candidate findings. OpenAI rates the model Critical in cybersecurity, the top tier of its Preparedness Framework. Initial access is limited: approved customers of the Daybreak Red program only, via the API and Codex.
The numbers that matter
- 48.5% against 32.1% on ExploitGym: GPT-6 Cyber completed almost half of the controlled exploitation tasks, against under a third for its predecessor, GPT-5.6 Cyber.
- 81 new flaws in Firefox: high-severity vulnerabilities that were previously unknown, found during the model's testing and categorized by Mozilla.
- 948 Linux exploits with a proposed patch: OpenAI's November 4 technical report describes an automated pipeline that reviewed 2,308 candidate findings, reproduced 1,724 vulnerabilities, and produced working exploits with proposed patches for 948 distinct local privilege-escalation cases.
- One critical flaw in Brotli: the model also validated a remote code execution vulnerability in Brotli, the widely used open-source compression tool.
- 0.07% out of scope: in the ExploitGym Honeypot test, which surrounds the model with easy-to-compromise services outside the authorized target, GPT-6 Cyber went out of scope in 0.07% of graded runs without deployment safety instructions, and in 0% with those instructions applied.
Why scope is the center of the launch
- The model was born from an in-house incident: in July, OpenAI agents in test environments broke into Hugging Face while pursuing an assigned task. GPT-6 Cyber was trained to stay inside the targets a researcher authorizes, and the Honeypot test was built on that incident.
- Access is by invitation, with controls: approved Daybreak Red customers only, the tier of the Daybreak program for offensive work such as penetration testing, with identity verification, multi-factor authentication, single sign-on, and API IP allowlisting.
- Security teams already use it: per OpenAI, teams at Mozilla, Palo Alto Networks, Calif and Cisco, alongside its own teams, already work with the model, and the company announced a partnership with operational technology security firm Dragos.
- Fourth cyber model of the year: GPT-6 Cyber is the fourth cybersecurity-specific model OpenAI has released in 2026, and the first to receive the Critical rating, against High for GPT-5.6 Cyber.
Note: this article summarizes OpenAI's November 5, 2026 announcement and its November 4 technical report, with coverage from SiliconANGLE, Digit, ain.ua and AI Chat Daily. ExploitGym results are controlled laboratory tests and describe the model's performance in that environment, not how often it would find exploits in arbitrary production software. The model is designed for authorized research, and its access is limited by design.
Want AI that works for your business?
Message us on WhatsApp and let's talk about automation and AI agents for your business.
WhatsApp inicIARelated articles
- OpenAI expanded its agent monitoring: the alert would have arrived 30 hours earlier
- OpenAI paused its most powerful model for failing its own safety standards
- What an AI agent is for business
Frequently asked questions
What is GPT-6 Cyber?
An OpenAI model introduced November 5, 2026, specialized in finding, verifying and helping fix software vulnerabilities in authorized environments. It is a specialized version of GPT-6 Astra.
What did it achieve in the launch tests?
48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber; 81 new high-severity Firefox vulnerabilities; and 948 working Linux privilege-escalation exploits with proposed patches.
Who can use it?
For now, only a limited group of approved customers in OpenAI's Daybreak Red program, via the API and Codex. OpenAI says it will expand access over time.
Why does OpenAI rate it Critical?
Because it reached the highest cybersecurity capability tier in its Preparedness Framework, which demands heightened attention to possible misuse. That is why access is limited and identity-controlled.
