News in Short

• GPT-5.6-Cyber completed 95 percent of an internal advanced cybersecurity test, according to OpenAI.
• OpenAI has introduced GPT-5.6-Cyber through its Daybreak Red access tier.
• The model is designed for advanced cybersecurity research and vulnerability discovery.
• OpenAI says it discovered two previously unknown vulnerabilities in Google’s V8 JavaScript engine.
• Trusted companies including SentinelOne and Palo Alto Networks received early access.

OpenAI is expanding its focus on cybersecurity with GPT-5.6-Cyber, a specialised model designed to help security researchers find and investigate vulnerabilities.

The model is being offered through the company’s Daybreak cybersecurity programme. Unlike its general-purpose models, GPT-5.6-Cyber is trained for advanced security work, including exploit development, vulnerability discovery and security research.

However, access is currently limited to selected companies and trusted defenders.

GPT-5.6-Cyber focuses on advanced security tasks

OpenAI says GPT-5.6-Cyber is designed to handle legitimate cybersecurity tasks with fewer unnecessary refusals.

The company tested the model across advanced scenarios. These included exploit-chain development, privilege escalation and vulnerability discovery.

According to OpenAI, GPT-5.6-Cyber completed 95 percent of requests in its internal evaluation. By comparison, GPT-5.6 Sol completed 1.5 percent, while GPT-5.6 Sol through Daybreak Blue completed 2 percent.

The company also says the new model performed better than GPT-5.5-Cyber. The earlier model completed 57.3 percent of the same requests.

These results suggest that OpenAI is treating cybersecurity as a specialised use case rather than simply adapting its general-purpose models.

Model reportedly found two unknown V8 vulnerabilities

One of the more notable tests involved V8, the JavaScript engine used by Google Chrome.

OpenAI says GPT-5.6-Cyber was used to investigate the software and discovered two previously unknown vulnerabilities. Researchers then validated the findings before reporting them to Google through coordinated disclosure.

Google subsequently fixed the vulnerability and assigned CVE-2026-15903.

The result highlights where specialised AI models could become useful for defensive security teams. Instead of waiting for vulnerabilities to be discovered by attackers or researchers, AI systems could help security teams examine large software projects for potential weaknesses.

OpenAI also tested GPT-5.6-Cyber on its Vulnerability Discovery and Report Writing evaluation.

Interestingly, GPT-5.6 Sol performed better than GPT-5.6-Cyber in this particular test. OpenAI attributed this partly to the Cyber model sometimes producing shorter and less detailed vulnerability reports.

Daybreak adds different levels of access

The GPT-5.6-Cyber launch is part of a broader expansion of OpenAI’s Daybreak cybersecurity programme.

The model is available through Daybreak Red. The company has also previously offered Daybreak Blue access, giving security researchers access to models with capabilities tailored for cybersecurity work.

OpenAI says GPT-5.6-Cyber also performed strongly on ExploitGym, an evaluation focused on advanced security tasks.

However, results varied depending on the amount of time given to the models. With the standard 300-turn limit on ExploitBench, GPT-5.6 Sol through Daybreak Blue performed best while using fewer tokens.

When the limit increased to 600 turns, the performance difference between GPT-5.6 Sol and GPT-5.6-Cyber became smaller.

This suggests that model performance can depend not only on the model itself, but also on how much time and computing resources an AI agent gets to complete a security task.

Security companies get early access

OpenAI has already provided GPT-5.6-Cyber to selected cybersecurity companies.

Early-access partners include SpecterOps, SentinelOne and Palo Alto Networks. The companies have used the model to improve defensive cybersecurity workflows.

The restricted access is significant because highly capable cybersecurity AI can have both defensive and offensive applications.

For OpenAI, the goal is to make advanced AI useful to security teams while controlling access to capabilities that could potentially be misused.

With GPT-5.6-Cyber, OpenAI is effectively moving deeper into a new phase of AI-assisted cybersecurity. The model is not simply designed to answer security questions. Instead, it is being positioned as a tool that can investigate software, identify vulnerabilities and assist researchers with complex security workflows.

The V8 findings also offer an early example of what such systems could contribute to real-world security. As AI models become better at reasoning through large codebases, their role in finding vulnerabilities before attackers do could become increasingly important.