OpenAI launches GPT-6 Cyber to automate vulnerability research

A new tool for deep code analysis

OpenAI officially announced GPT-6 Cyber on November 5, 2026. This model is specifically designed for cybersecurity tasks such as vulnerability research, exploit validation, and patch development. It goes beyond basic code scanners by looking at how different parts of a program interact with one another.

Traditional security tools often flag simple patterns in code. However, many serious weaknesses sit deep within complex software where these scans rarely reach. GPT-6 Cyber is meant to fill this gap by checking whether a possible weakness actually poses a real security risk to the user.

As Analytics Insight reported, the model can investigate various risks including SQL injection, cross-site scripting, and memory errors. It can also test whether a suspected flaw allows someone to enter a system without authorization or leak sensitive information. This provides developers with firmer evidence when deciding which bugs to fix first.

Proven results in browser and kernel security

The model's capabilities were demonstrated through research involving Mozilla Firefox. OpenAI reported that the AI found 81 previously unknown high-severity vulnerabilities in the browser. These results are significant because popular applications run on millions of lines of code that are difficult for humans to audit completely.

In addition to the Firefox findings, the research produced 948 working exploits for Linux kernel vulnerabilities. The model also generated candidate patches for these issues. These results show that AI can help researchers comb through massive codebases to find suspicious behaviors that might otherwise stay hidden.

While these numbers are impressive, the developers emphasize that human review remains essential. Security experts must verify every finding and confirm the final solution before any official approval. The goal is to provide researchers with more ways to probe code rather than replacing the human oversight entirely.

How the model assists the dev cycle

Finding a vulnerability is only the first step in the security process. Developers also need to understand the root cause of a flaw and ensure that a fix does not break other parts of the application. GPT-6 Cyber can explain the weakness and suggest specific code changes to resolve it.

For example, if an API lets users reach data without proper permission checks, the AI can trace the request path to spot the missing check. It can then suggest a safer approach for the developer to implement. This process can shorten the time between discovering a bug and repairing it in a production environment.

The model can even write tests to check whether a proposed fix actually works. However, developers must be cautious because AI-written patches might introduce new bugs. The model is intended to be used alongside established tools like Semgrep, GitHub CodeQL, and Snyk to provide a multi-layered defense.

Controlled access and performance

OpenAI is not making GPT-6 Cyber available to the general public. Instead, it ties advanced cybersecurity work to a controlled access system called the Daybreak program. This program includes a Blue track for defensive work like malware analysis and a Red track for more advanced, authorized security research.

The model features a capacity of 1,050,000 tokens for its context window and can produce up to 128,000 tokens in a single output. Pricing is set at USD 20 per million input tokens and USD 100 per million output tokens. These figures indicate the scale of the model, though specific usage limits and approval processes still apply to users.

Performance benchmarks show a significant jump in capability for security tasks. The model reached a 48.5% score on the ExploitGym evaluation, which is a relative improvement of about 51% over the previous GPT-5.6 Cyber model. It also reached 40.3% recall in web vulnerability discovery, suggesting a real step up in handling hard security tasks.

Limitations and the human element

Despite its capabilities, AI tools can still throw false alarms or miss hidden flaws. A study on Vul-RAG from late 2026 indicated that language models struggled to tell vulnerable code from similar patched code in many tested cases. This means users should not trust AI output without independent verification.

The best results come from pairing AI research with careful human judgment. While the model can speed up the discovery of flaws and improve patch quality, it cannot replace the expert's ability to understand the broader context of a system. It serves as a powerful assistant to help protect user data and cut the risk of costly attacks.

Ultimately, better security depends on whether the final fix protects the software without opening new weaknesses. By using GPT-6 Cyber to investigate vulnerabilities and draft fixes, developers can move faster while maintaining a high standard of safety. The focus remains on using AI to augment the existing security process rather than automating it entirely.