OpenAI has unveiled Daybreak, a security initiative designed to find and fix code vulnerabilities before attackers can exploit them. The system combines the company’s Codex Security agent, released in March, with its latest cyber-focused models to automate threat detection and patching.

Daybreak works by analyzing an organization’s codebase to build a threat model, mapping out possible attack paths. It then validates which vulnerabilities are most likely to be exploitable and automates the detection of high-risk issues. The system relies on a mix of models, including GPT-5.5-Cyber and GPT-5.5 with Trusted Access for Cyber, which began rolling out last week.

“We’re seeing a shift from reactive patching to proactive defense,” said a senior security engineer at a financial firm that tested the tool, speaking on condition of anonymity. “Instead of waiting for a breach, Daybreak surfaces the problems that matter most.”

The launch comes roughly a month after rival Anthropic quietly released Claude Mythos, a security model it deemed too dangerous for public release and shared only privately under its Project Glasswing. Despite restrictions, some unauthorized users gained access.

OpenAI had lacked a comparable security offering until now. Like Glasswing, Daybreak is not a single model but a system that pulls together multiple AI capabilities and partner integrations. OpenAI says it is working with industry and government partners as it prepares to deploy increasingly capable cyber models.

Source: The Verge