OpenAI has introduced a new cybersecurity initiative, Daybreak, aimed at enhancing software security. This launch directly competes with Anthropic's Project Glasswing, which leverages the unreleased AI model, Claude Mythos Preview, to bolster clients' cyber defenses. Mozilla reported that the Mythos model effectively identified and addressed 271 vulnerabilities in the Firefox browser in April.
Daybreak utilizes several AI models, particularly its security agent Codex, to proactively integrate cyber defense measures into software. The initiative focuses on tackling significant security issues, streamlining the process from hours of analysis to mere minutes, and ensuring audit-ready results for clients. For instance, Codex Security was tasked with scanning a codebase to identify and resolve the most critical findings.
The initiative will employ GPT-5.5 for general functions and a specialized version, GPT-5.5 with Trusted Access for Cyber, for various defensive tasks, including secure code review and malware analysis. Additionally, GPT-5.5-Cyber will be used for specialized operations like penetration testing. OpenAI is collaborating with multiple partners, such as Cloudflare, Cisco, Palo Alto Networks, and others, to advance Daybreak's capabilities.