Openai Cyber Model Expands Daybreak Initiative With 2 New Tiers

In the past few weeks, advanced AI models have increasingly shown unexpected, potentially malicious behaviour such as breaching the security infrastructure of companies and engaging in deception by creating fake profiles. Amid the recent surge in cybersecurity incidents, OpenAI on Monday, August 10, introduced its new cybersecurity-focused AI model and also announced an expansion of its cybersecurity initiative, Daybreak, that offers limited access to models, tools, and workflows for cyber defenders.

The use of AI by bad actors is speeding up cyber attacks at an unprecedented pace, leaving virtually no time for defenders to prepare and act. OpenAI believes that frontier intelligence in the hands of trusted defenders is the best way to deploy offensive AI capabilities at scale. As part of its efforts in this direction, the Sam Altman-led company is offering OpenAI Daybreak with two access tiers to provide the right set of capabilities to the trusted defenders.

The ChatGPT-maker also announced GPT-5.6-Cyber, which is available through Daybreak Red. The new cyber model which is backed by GPT-5.6 Sol has been trained to improve capabilities on several specialised cybersecurity tasks, such as finding zero-day vulnerabilities and developing exploit chains, and to reduce refusals for certain high-risk tasks. At present, GPT-5.6 Cyber is only available for ‘trusted customer partners’, which reportedly include Crowdstrike, IBM, Cloudflare, etc.

Daybreak unlocks advanced cyber capabilities. Earlier, GPT-5.6 Sol, which delivered state-of-the-art performance on cybersecurity tasks, owing to its safeguard, could also block legitimate defensive work. OpenAI said that Daybreak Blue access removes those guardrails, essentially helping defenders with real-world security tasks like incident detection and response, investigations, vulnerability management, and security assessments.

Last month, a number of AI-related cybersecurity incidents were reported with the most notable being OpenAI’s frontier models escaping sandboxed environments during cybersecurity testing. The models breached Hugging Face’s production infrastructure in their effort to accomplish a cybersecurity capability benchmark. On July 16, Hugging Face publicly disclosed that the autonomous AI system has breached its infrastructure. Two days later, OpenAI identified the escalation path in its own systems and linked the breach to its AI agent.