OpenAI is partnering with major global security companies to expand the rollout of 'GPT-5.6 Cyber,' an artificial intelligence (AI) model specialized in cybersecurity. The goal is to provide customized security AI to organizations and professionals engaged in defense operations.

On the 11th, OpenAI announced the expansion of the 'Daybreak Cyber Partner' program with leading global security firms, including Accenture, Cisco, Cloudflare, CrowdStrike, IBM, and Palo Alto Networks Unit 42. Through this partnership, each company can integrate OpenAI's advanced cybersecurity models into their own security products, managed security services (MSS), and customer projects. Customers who have completed Amazon Bedrock registration can also access OpenAI's Daybreak Blue and Red services.

'GPT-5.6 Cyber' and Customized 'Daybreak' Unveiled

The newly introduced 'GPT-5.6 Cyber' is an AI model specialized for advanced vulnerability research, exploit verification, and security testing. In conjunction with the model's release, OpenAI also unveiled the 'Daybreak' program, which provides customized security AI to individuals and organizations conducting authorized defense operations.

Daybreak is divided into two categories based on intended use, and only pre-approved users can access them.

• Daybreak Blue: Designed for general defense missions, based on 'GPT-5.6 Sol.' OpenAI recommended that most security teams use this module as a starting point.
• Daybreak Red: Directly provides 'GPT-5.6 Cyber' to organizations requiring professional security capabilities, such as advanced vulnerability detection, exploit development, and red team penetration testing.

Both environments are subject to strict multi-layered security controls, including identity verification, account security, real-time monitoring, restrictions on use beyond approved purposes, and legal compliance verification.

Top-Tier Benchmark Performance and Real-World Vulnerability Discoveries

'GPT-5.6 Cyber' recorded the highest success rate in the 'ExploitGym' benchmark among all models evaluated by OpenAI. This benchmark assesses the ability to convert known vulnerabilities into code that can be used in actual attacks. Additionally, it demonstrated outstanding performance by successfully completing 95% of all requests in an evaluation of high-difficulty cybersecurity task execution rates.

Remarkable results were also achieved in real-world software vulnerability research. During the analysis of Google Chrome's JavaScript engine 'V8,' OpenAI researchers discovered two previously unreported vulnerabilities that could be chained together for exploitation and reported them to Google, which promptly applied security patches for the vulnerabilities.

Furthermore, in a recent cyber evaluation, 'GPT-5.6 Sol' and a research model accessed the actual Hugging Face system and discovered new vulnerabilities. They also secured attack paths enabling privilege escalation and remote code execution, confirming that AI's offensive capabilities are rapidly evolving.

'Defender's Advantage' Principle and Emphasis on Industry Collaboration

OpenAI stated that as AI's offensive capabilities become more sophisticated, the same technology must be proactively supplied to defenders. Accordingly, the company adopted a dual-track approach: providing 'GPT-5.6 Sol' with safety guardrails for general defense operations, and providing 'GPT-5.6 Cyber' for advanced exploit development only after a separate review process.

Clem Delangue, Co-Founder and CEO of Hugging Face, emphasized that "AI safety cannot be solved by one company developing it behind closed doors; security professionals worldwide must broadly leverage AI and collaborate."