OpenAI Introduces GPT-5.6-Cyber and Expands Cybersecurity Program

OpenAI is launching GPT-5.6-Cyber, expanding its Daybreak cybersecurity program to two access tiers, and simultaneously growing its Daybreak partner program. Participating security companies, consulting firms, and technology providers include IBM, Accenture, Cisco, CrowdStrike, Palo Alto Networks, Cloudflare, PwC, and KPMG, among others. Partners can integrate OpenAI's cyber models into their own products, services, and customer projects.

Daybreak was launched in June as an umbrella program for OpenAI's cybersecurity offerings, providing authorized individuals and organizations with controlled access to highly capable AI models for cybersecurity tasks.

In the future, there will be two access tiers: Daybreak Blue offers general top-tier models like GPT-5.6 Sol with reduced system-level cyber safeguards for authorized security work. Daybreak Red, on the other hand, is for more demanding tasks such as vulnerability research, exploit validation, and security testing, granting access to models specifically trained for cybersecurity. This includes the new GPT-5.6-Cyber, exclusively available through Daybreak Red, and is the third model specifically tailored for cybersecurity, following GPT-5.4-Cyber and GPT-5.5-Cyber.

GPT-5.6-Cyber Shows Strengths in Exploits and Zero-Days

GPT-5.6-Cyber is based on GPT-5.6 Sol but has been specifically trained for demanding cybersecurity tasks, particularly the search for previously unknown zero-day vulnerabilities and the development of exploit chains. While Daybreak Blue loosens system-level safeguards for GPT-5.6 Sol for authorized security work, GPT-5.6-Cyber has also been trained to refuse cooperation significantly less often, even for highly sensitive dual-use tasks (usable defensively and offensively).

GPT-5.6-Cyber also demonstrates advantages in several cybersecurity benchmarks. In ExploitGym, which tests the development of functional exploits for known vulnerabilities in isolated test environments, it outperforms GPT-5.6 Sol and GPT-5.5-Cyber. In an internal test for discovering and evaluating previously unknown zero-day vulnerabilities, the new model also leads GPT-5.6 Sol.

OpenAI has also applied GPT-5.6-Cyber to real-world software. In Google's JavaScript engine V8, the model discovered two previously unknown vulnerabilities that, according to OpenAI, could be combined to manipulate memory and break out of the V8 heap sandbox. The findings were reported to Google, and one of the vulnerabilities has already been fixed. Furthermore, OpenAI claims to have found at least five vulnerabilities in a widespread mobile operating system, three critical vulnerabilities in a common database, and over 400 privilege escalation vulnerabilities in a common operating system kernel. The company has not disclosed the affected products.

After Hugging Face Incident: OpenAI Tightens Security Measures

Access to Daybreak Blue and Red remains restricted to verified individuals and organizations with authorized security tasks. OpenAI relies on identity verification, monitoring, and additional account security measures, and intends to further tighten these safeguards.

These measures come shortly after an incident where OpenAI models autonomously compromised Hugging Face systems during security tests. According to OpenAI, GPT-5.6-Cyber was not involved. Since then, the company has tightened its security controls for internal tests and has temporarily suspended certain activities with the unreleased model Astra, for which it cannot rule out critical cyber capabilities. temporarily suspended.

(wpl)