OpenAI Launches GPT-6 Cyber for Security Researchers With Critical Cybersecurity Rating

OpenAI has launched GPT-6 Cyber, a specialised artificial intelligence model designed to help cybersecurity researchers identify software vulnerabilities, develop working exploits and strengthen digital infrastructure. Announced on November 5, 2026, the model represents a major expansion of OpenAI’s cybersecurity capabilities and has received a Critical cybersecurity classification under the company’s Preparedness Framework.

The model is designed to support advanced security research while following the boundaries established by authorised researchers. OpenAI says GPT-6 Cyber has already helped identify 81 previously unknown high-severity vulnerabilities in Mozilla Firefox and develop working exploits for hundreds of Linux kernel vulnerabilities. However, access is initially restricted to a select group of approved security organisations because the same capabilities could potentially be misused by malicious actors.

What Is OpenAI’s GPT-6 Cyber?

GPT-6 Cyber is a cybersecurity-focused version of OpenAI’s GPT-6 Astra model. It has received additional training to perform specialised security tasks, including vulnerability discovery, exploit development, penetration testing and security validation.

Unlike general-purpose AI assistants, which are primarily designed to answer questions, write code and assist with everyday tasks, GPT-6 Cyber is optimised for complex security investigations.

Its capabilities include:

• Identifying weaknesses in software and applications.
• Reproducing vulnerabilities to establish whether they are exploitable.
• Developing proof-of-concept exploits.
• Testing security fixes and patches.
• Analysing large software codebases.
• Supporting authorised penetration testing and red-team operations.

The model can help security researchers investigate vulnerabilities that may be difficult to identify through conventional automated scanning tools.

For example, a researcher can use it to investigate a complicated software component, identify a potential weakness, reproduce the issue and develop a technical demonstration that helps engineers understand the problem.

OpenAI says the model outperforms its general-purpose GPT-6 Astra system on most cybersecurity tasks.

Why Has OpenAI Given GPT-6 Cyber a Critical Rating?

One of the most important aspects of the launch is the model’s cybersecurity classification.

OpenAI has assigned GPT-6 Cyber a Critical rating under its Preparedness Framework. Its predecessor, GPT-5.6 Cyber, received a High classification.

The Critical designation reflects the model’s ability to identify vulnerabilities and develop working exploits at a level that could create significant cybersecurity risks if misused.

It is important to understand that this is a risk classification, not a conventional quality or performance rating. It indicates the potential consequences of the model’s capabilities rather than simply describing how good it is.

OpenAI has therefore introduced additional restrictions around access, monitoring and deployment.

GPT-6 Cyber Scores 48.5% on ExploitGym

OpenAI reported substantial improvements in its internal and specialised cybersecurity evaluations.

On ExploitGym, a benchmark that tests whether AI models can exploit specified vulnerabilities in remote targets, GPT-6 Cyber achieved a score of 48.5%.

Its predecessor, GPT-5.6 Cyber, scored 32.1%.

Model	ExploitGym score
GPT-6 Cyber	48.5%
GPT-5.6 Cyber	32.1%

This represents a 16.4-percentage-point improvement, equivalent to approximately 51% relative growth over the previous model.

OpenAI also tested the model on web applications without giving it access to their source code. GPT-6 Cyber achieved 40.3% vulnerability recall, compared with 32% for GPT-6 Astra.

Another evaluation, RemoteExploitBench, tested models against controlled deployments of applications including GitLab, PostgreSQL and Jenkins. GPT-6 Cyber completed 95% of the tasks in the initial evaluation, while GPT-6 Astra refused all the tasks under the tested configuration.

These results demonstrate improvements in the model’s ability to conduct practical vulnerability research. However, benchmark scores should not be interpreted as a guarantee that it can successfully compromise real-world systems.

GPT-6 Cyber Discovers 81 Firefox Vulnerabilities

One of the most significant demonstrations involved Mozilla Firefox, the widely used open-source web browser.

OpenAI researchers used GPT-6 Cyber to investigate Firefox and identify 81 previously unknown vulnerabilities that Mozilla classified as high severity.

The model also identified 30 previously reported issues that were more serious than originally assessed.

OpenAI provided Mozilla with evidence, reproducible tests and proposed patches to help its engineering teams investigate and address the vulnerabilities. Mozilla subsequently began releasing fixes for many of the reported issues.

The research is notable because Firefox is a heavily scrutinised software project that has undergone years of security testing.

The results suggest that advanced AI systems could help researchers discover weaknesses that might otherwise require substantial manual investigation.

However, the findings also demonstrate why powerful vulnerability-discovery models need carefully controlled access. The ability to discover previously unknown vulnerabilities can help software developers improve security, but similar capabilities could be used to identify weaknesses before they are patched.

AI Model Also Helps Discover Linux Kernel Exploits

OpenAI has also reported significant results from research involving the Linux kernel, which powers a large proportion of the world’s servers, cloud infrastructure and computing devices.

According to the company, GPT-6 Cyber helped researchers reproduce 1,724 Linux kernel vulnerabilities and develop working exploits for 948 of them.

OpenAI also disclosed 304 critical findings to maintainers, along with candidate patches.

These results demonstrate the potential of specialised AI systems to assist with large-scale vulnerability research. Instead of manually investigating every possible weakness, security researchers can use AI to examine software, test potential attack paths and prepare reproducible evidence.

For maintainers, this could make it easier to prioritise security fixes and understand how vulnerabilities might affect downstream applications.

The Linux findings are particularly relevant because weaknesses in widely deployed operating-system components can have consequences across cloud services, enterprise infrastructure and connected devices.

OpenAI Focuses on Preventing Unauthorised Cyberattacks

The capabilities of GPT-6 Cyber create a difficult challenge for OpenAI.

A model that can identify vulnerabilities and develop working exploits can be valuable to legitimate security researchers. The same model could also assist attackers attempting to compromise systems without permission.

OpenAI says it has therefore placed significant emphasis on ensuring that GPT-6 Cyber respects authorised testing boundaries.

In its ExploitGym Honeypot evaluation, researchers created isolated environments containing supporting services that were deliberately vulnerable. Exploiting those services could help complete a task, but doing so would violate the authorised scope.

Without deployment safety prompts, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of evaluated cases. With the additional safety prompts enabled, it stayed within the authorised boundaries in 100% of the tests.

For comparison, GPT-5.6 Cyber achieved 94.49% without the additional prompts.

OpenAI is also using cyber-specific monitoring, credential protections, IP allowlisting, sandboxing recommendations and human oversight for high-risk workflows.

The company has introduced additional safeguards for critical infrastructure, including systems involved in electricity generation, water treatment and industrial control operations.

Who Can Access GPT-6 Cyber?

GPT-6 Cyber is not being released as a publicly available chatbot.

Initial access is limited to selected, approved customers through OpenAI’s Daybreak Red programme, which supports advanced authorised cybersecurity work.

The model is available through the API and Codex for these approved customers. OpenAI plans to expand access to other Daybreak Red participants.

The company is also making GPT-5.6 Cyber available to all Daybreak Red customers and expanding access to Daybreak Blue, which supports verified defenders using GPT-6 Astra for security validation and defensive research.

This restricted rollout reflects the potential dual-use risks associated with advanced vulnerability research.

The Bigger Picture

GPT-6 Cyber highlights how artificial intelligence is changing the cybersecurity industry. Traditional security teams often rely on specialised tools, experienced researchers and lengthy manual investigations to identify complex software vulnerabilities.

Advanced AI models could accelerate parts of this process by analysing code, reproducing weaknesses and helping developers prepare patches.

At the same time, the technology creates new risks. As AI systems become better at developing working exploits, attackers could potentially use similar tools to reduce the time and expertise required for sophisticated cyber operations.

The challenge for AI developers will be ensuring that these capabilities primarily benefit defenders. OpenAI’s restricted access model and additional safeguards represent one approach to managing that risk, although no security mechanism can guarantee that misuse will never occur.

Looking Ahead

The next stage for GPT-6 Cyber will depend on how effectively OpenAI expands access while maintaining strong security controls. Wider adoption among software companies, open-source maintainers and authorised penetration-testing teams could help identify vulnerabilities earlier and improve the speed of security patching. Independent assessments will also be important for understanding how reliably the model respects authorisation boundaries outside controlled testing environments.

For the broader AI industry, the launch signals that specialised cybersecurity models are becoming increasingly capable of performing advanced technical research. Their impact will depend not only on how many vulnerabilities they can discover, but also on how responsibly those capabilities are deployed. GPT-6 Cyber could become an important tool for digital defence, but its powerful exploit-development capabilities make access control and oversight essential.