OpenAI's GPT-6 Cyber Found 81 Unknown Firefox Flaws and Built 948 Linux Exploits

OpenAI's new cybersecurity model, GPT-6 Cyber, found 81 previously unknown high-severity vulnerabilities in Firefox and produced working exploits for 948 Linux kernel flaws, 588 of them never documented before. The company announced the model on November 5 and is restricting access to a small group of approved security customers because the same capabilities that find flaws can also attack systems.

OpenAI says GPT-6 Cyber outperforms its general-purpose GPT-6 Astra model on most cyber tasks. Fortune first reported in September that OpenAI planned to release a specialized cyber model built on GPT-6.

Why does a security model need restricted access?

The model focuses on advanced vulnerability research and exploit development. To test a server's defenses, researchers break into it to see what an attacker could do with a flaw. The same methods power malicious attacks, so general-purpose AI models impose restrictions that also block legitimate security work. OpenAI said it trains its Cyber models both to sharpen specialized cybersecurity skills and to reduce refusals for tasks that serve attackers and defenders alike.

The test targets explain the stakes. Firefox must handle potentially malicious webpages without exposing user data. The Linux kernel controls access to a machine's resources. Both are open-source projects, inspected and hardened by developers and security researchers for years. AI models still find novel vulnerabilities in them.

The Linux results show how fast the frontier is moving. The new exploits could let an attacker with ordinary local access gain root privileges, the highest level of control over a system. Researchers developed working exploits for more than five times as many Linux vulnerabilities as in a July study using GPT-5.6 Sol, which was considered state-of-the-art for cyber tasks when released that month.

Who gets to use it?

OpenAI runs two access tiers for trusted cybersecurity professionals through its Daybreak program:

• Daybreak Red supports offensive-style work: simulating attacks and building red-teaming tools.
• Daybreak Blue covers defensive tasks: reviewing code, analyzing malware and investigating incidents.

GPT-6 Cyber will initially reach a select group of approved Daybreak Red customers only. Access expands to all Red customers over the coming weeks, the company said. Separately, OpenAI is widening early testing of Astra for defensive security work to more organizations and individual researchers through Daybreak Blue.

How big is the market OpenAI is chasing?

Gartner forecast in September that worldwide AI cybersecurity spending will reach $51.3 billion this year, nearly double last year's $25.9 billion. As AI models get better at finding software flaws and turning them into working attacks, businesses face increasing pressure to strengthen their defenses.

OpenAI is leaning into that demand. GPT-6 Cyber is the company's fourth cybersecurity-focused model this year, following releases in April, June and August. The company also hired a chief revenue officer, Dali Rajic, who comes from a cybersecurity software background. OpenAI views security as an important path to growing enterprise sales.

Where do Google and Anthropic stand?

OpenAI is not alone in courting security customers. Anthropic's Glasswing program gives vetted users access to capabilities that its public models might block. Google previously emphasized a lightweight, low-cost model for repeatedly scanning large codebases. It now also offers vetted defenders its frontier Gemini 4 Argon model without cyber guardrails through its Fairwind Program.

The pattern across all three vendors is the same: frontier models stripped of refusal behavior, released only to vetted defenders, and benchmarked on real targets. With AI-driven attacks rising and budgets nearly doubling in a year, the competition for trusted-security revenue is likely to tighten further — and the guardrails will keep living in the customer vetting, not in the models.