The model has identified 81 previously unknown high-severity Firefox flaws and developed exploits for 948 Linux kernel vulnerabilities.
OpenAI has introduced GPT-6 Cyber, a specialized AI model designed to discover software vulnerabilities, develop working exploits and help security teams validate fixes, as the company seeks to expand the use of advanced AI in cybersecurity while limiting its potential misuse.
Announced November 5, GPT-6 Cyber is designed to perform complex security research, including identifying weaknesses in large codebases and developing proof-of-concept exploits. OpenAI said the model also makes significant improvements in staying within the boundaries of authorized security testing, addressing a key concern as AI systems become increasingly capable of offensive security tasks.
The company said GPT-6 Cyber achieved a 48.5% score on ExploitGym, a benchmark that evaluates a model’s ability to exploit specified vulnerabilities in remote targets. Its predecessor, GPT-5.6 Cyber, scored 32.1%, representing a 51% relative improvement.
OpenAI also reported that the model achieved 40.3% recall in a new benchmark for identifying vulnerabilities in web applications without access to their source code, compared with 32% for GPT-6 Astra, its general-purpose model.
The improvements extend beyond benchmark performance. OpenAI said its researchers used GPT-6 Cyber to identify 81 previously unknown high-severity vulnerabilities in Firefox and develop working exploits for 948 Linux kernel vulnerabilities.
The Firefox findings were reported to Mozilla, along with proposed patches. Mozilla has already released fixes for many of the identified vulnerabilities. In the Linux kernel research, the model reproduced 1,724 vulnerabilities and developed working exploits for 948 of them. OpenAI said it disclosed 304 critical findings to Linux maintainers, along with candidate patches, while work continues on other vulnerabilities.
The findings illustrate how AI models are moving beyond assisting developers with code analysis to performing substantial portions of vulnerability research and exploit validation.
Keeping AI security research within scope
A central focus of GPT-6 Cyber’s development has been ensuring that the model does not extend its activities beyond the systems and services it has been authorized to test.
An AI agent conducting penetration testing could potentially discover and exploit weaknesses in connected services that are not part of the assigned engagement. Such behavior can create additional risks, even when the original task is legitimate security research.
OpenAI said it introduced specialized training to make GPT-6 Cyber more reliable at following restrictions, respecting authorized testing boundaries and accurately reporting its findings.
In an evaluation called ExploitGym Honeypot, the company tested whether the model would exploit deliberately vulnerable supporting services that were outside the scope of its assigned task. Without deployment safety prompts, GPT-6 Cyber avoided exploiting these services in 99.93% of tests, compared with 94.49% for GPT-5.6 Cyber. With deployment prompts enabled, it stayed within the authorized scope in all tests.
However, OpenAI also classified GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. The classification reflects the model’s increased ability to identify vulnerabilities and develop working exploits.
The company has introduced additional safeguards, including cyber-specific monitoring to detect activities that deviate from user authorization, automated checks for potentially malicious use and restrictions on access to its more advanced exploit-development capabilities.
Critical infrastructure remains a concern
OpenAI is also addressing the implications of advanced AI-assisted security research for critical infrastructure, including power plants, water treatment facilities and industrial control systems.
The company worked with operational technology cybersecurity provider Dragos to evaluate how its models could interact with industrial environments. The assessment involved a simulated power plant and tested activities ranging from network discovery and asset identification to sensitive data extraction and potentially dangerous device operations.
According to OpenAI, the evaluation identified two previously unknown vulnerabilities in engineering software responsible for managing protective relay data. Such software can expose sensitive information about protection settings and circuit-breaker operations.
The exercise was conducted in a controlled laboratory environment. OpenAI said feedback from Dragos helped refine safeguards designed to prevent assistance that could compromise industrial controls or cause physical damage, while preserving legitimate security research.
The company is also introducing a dedicated Daybreak Red for Critical Infrastructure access group for approved defenders. This will allow authorized testing of critical systems under additional safeguards.
OpenAI said its $1 billion Daybreak for Frontline Defenders grant program will subsidize access to advanced security capabilities for resource-constrained organizations, including government bodies and essential service providers.
Access remains restricted
GPT-6 Cyber is initially available through the API and Codex to a select group of approved Daybreak Red customers. OpenAI plans to expand access to other Daybreak Red customers progressively.
The company is also expanding access to Daybreak Blue, which is designed to help verified defenders validate vulnerabilities and fixes with fewer unnecessary restrictions. Daybreak Red supports more advanced exploit development and authorized security testing using GPT-6 Cyber.
The distinction reflects OpenAI’s approach to providing different levels of cybersecurity capabilities depending on the user’s authorization and intended work.
OpenAI said security teams at Cisco, Calif and CodeRabbit are already using GPT-6 Cyber for vulnerability research and code review. The company has also used the model to identify and address weaknesses in its own infrastructure.
As AI systems become more capable of discovering and exploiting software flaws, the challenge for security teams will increasingly involve managing both sides of the technology: using AI to accelerate vulnerability discovery and remediation while ensuring that similar capabilities do not lower the barriers for malicious exploitation.
