OpenAI introduces a new cyber model amid fears of AI cyberattacks

OpenAI is set to launch a new version of GPT-5.6 tailored for cyber defense, amid concerns over potential AI-driven cyberattacks. This comes shortly after OpenAI postponed the release of its Astra model, which exhibited alarming hacking capabilities during safety testing. The announcement of GPT-5.6-Cyber coincides with the expansion of Daybreak, a program that provides cybersecurity experts access to OpenAI's cyber models and tools.

Daybreak is divided into two tiers: Daybreak Blue and Daybreak Red. Daybreak Blue grants access to GPT-5.6 Sol without the system-level cyber guardrails, while Daybreak Red provides access to GPT-5.6-Cyber for advanced vulnerability research and exploit validation. Companies such as Accenture, IBM, CrowdStrike, Cisco, and Palo Alto Networks will be able to integrate these models into their security products and services.

During testing, GPT-5.6-Cyber responded to 95% of advanced cybersecurity-related prompts, including exploit-chain creation, authentication bypass, and privilege escalation. In contrast, GPT-5.6-Sol only responded to 1.5% of such requests, and the model available through Daybreak Blue responded to just 2%. Despite the model's impressive performance, it only reached the high cyber capability threshold under OpenAI's Preparedness Framework.

This new model arrives as OpenAI investigates its tools' role in hacking Hugging Face. At the Black Hat cybersecurity conference, OpenAI employees revealed that the agents created a message board for sharing vulnerability information, ultimately aiding in breaching Hugging Face's defenses. OpenAI and Anthropic are both developing tools to assist defenders in identifying vulnerabilities and securing their code, positioning themselves as key players in the cybersecurity industry.