OpenAI Launches GPT-6 Cyber With 51% ExploitGym Gain

By Amara Osei

OpenAI said GPT-6 Cyber lifted ExploitGym success to 48.5% from 32.1% for GPT-5.6 Cyber, a 51% improvement announced on November 5, 2026. OpenAI uses ExploitGym to test whether models can exploit specified vulnerabilities in controlled environments.

What performance did OpenAI report?

OpenAI's web-penetration evaluation, run without source-code access, found 40.3% of vulnerabilities with GPT-6 Cyber, compared with 32% for GPT-6 Astra. OpenAI also reported that GPT-6 Cyber completed 95% of RemoteExploitBench tasks involving known vulnerabilities in GitLab, PostgreSQL and Jenkins.

OpenAI classified GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework, while GPT-5.6 Cyber received a High classification. The launch expands OpenAI's development of AI for cybersecurity research, vulnerability discovery and defensive security operations, the company said.

How did OpenAI test misuse controls?

OpenAI said strengthening alignment and preventing unauthorized security activity was central to GPT-6 Cyber's development. The company reported that the model avoided exploiting out-of-scope services in 99.93% of ExploitGym Honeypot tests without deployment-specific safety instructions, versus 94.49% for GPT-5.6 Cyber.

With those instructions enabled, GPT-6 Cyber stayed within authorized boundaries in 100% of evaluated cases. OpenAI said the results are intended to let security professionals use offensive research capabilities without letting AI agents expand beyond permitted systems.

What did the model find in Firefox?

OpenAI said GPT-6 Cyber identified 81 previously unknown high-severity vulnerabilities during an investigation of the Firefox browser; Mozilla classified all 81 as high severity. The model also produced evidence that another 30 previously identified issues were more severe than initially reported. OpenAI shared the findings and candidate patches with Mozilla, enabling engineers to begin addressing the vulnerabilities.

Brian Grinstead, Distinguished Engineer, Firefox, said:

“GPT‑6 Cyber is a world-class model for cyberdefense. Firefox entered 2026 as one of the world’s most hardened and scrutinized codebases.

Even after further hardening with multiple generations of frontier models, GPT‑6 Cyber uncovered previously unknown high-severity vulnerabilities, identified architectural improvements, and helped us ship fixes to millions of users.”

What did the Linux research show?

In separate Linux kernel research, OpenAI said GPT-6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948. Under the same research framework, GPT-5.6 Sol successfully exploited 166 vulnerabilities.

OpenAI disclosed 304 critical Linux findings to maintainers and provided potential fixes through coordinated disclosure. The company also used GPT-6 Cyber with other models to investigate Brotli, a compression library used by Node.js and other applications.

OpenAI said security teams at Cisco, Calif and CodeRabbit are using the model for vulnerability research and code review. The company also uses GPT-6 Cyber in defensive assessments of its own infrastructure, including investigations involving isolated computing environments.

Can the model work on critical infrastructure?

OpenAI partnered with industrial cybersecurity company Dragos to evaluate AI models in simulated operational technology environments. During controlled testing of a mock power plant, a frontier model identified operational technology assets and found two previously unknown vulnerabilities in engineering software used to manage protective relay data.

Rob T. Lee, Chief AI Officer, Dragos, said:

“Dragos put a frontier model on a mock power plant range and told it almost nothing. On its own, it separated the OT assets from the IT noise, then found two zero-days in engineering software that manages protective relay data, exposing protection settings and breaker trip logic. That’s the map an adversary wants before touching anything physical.

The skill floor is dropping for the reconnaissance and vulnerability research needed before an attack. Working with OpenAI on this shows them the adversary’s path through an OT network before real adversaries take it, so safeguards get built around how these attacks could unfold, while the models stay useful to the people defending these systems.”

OpenAI said the mock-plant findings showed both the defensive value of advanced AI research and the risks of increasingly capable autonomous cybersecurity systems. Feedback from Dragos informed safeguards for industrial control systems, including protections against unauthorized activity affecting power plants, water treatment facilities and other safety-critical environments.

How is access being controlled?

OpenAI said its Daybreak cybersecurity program grants different levels of model access according to defenders' needs and authorization:

• Daybreak Blue supports verified defenders conducting vulnerability research and validation.
• Daybreak Red provides more advanced capabilities for authorized penetration testing and exploit development.
• Daybreak Red for Critical Infrastructure is a specialized category for approved organizations protecting industrial and essential service environments.

The company said its $1 billion Daybreak for Frontline Defenders grant program will subsidize access to advanced security technology for resource-constrained organizations, including government agencies and essential service providers. OpenAI also listed four accompanying safeguards:

• Cybersecurity-specific misalignment monitoring
• Threat detection
• Stronger credential protections
• Enhanced oversight of critical-infrastructure activities

OpenAI initially made GPT-6 Cyber available through its API and Codex to a select group of approved Daybreak Red customers, with plans to expand access. The company is also broadening access to GPT-5.6 Cyber and GPT-6 Astra through existing cybersecurity programs.

OpenAI is positioning GPT-6 Cyber to accelerate vulnerability discovery, validate security weaknesses and strengthen critical software as AI-generated software becomes more widespread and cyber threats grow more sophisticated. As access expands, OpenAI must preserve the 100% authorized-boundary result it recorded with deployment safety instructions.