OpenAI on July 12 introduced Daybreak, a cybersecurity AI program designed to detect and help fix system weaknesses before attackers find them. The offering is built around the Codex Security AI agent launched in March, which can model threats across enterprise codebases, map likely attack paths, and automate detection workflows for high-risk vulnerabilities.
The release also puts OpenAI more directly into competition with Anthropic’s security push. While Anthropic’s Claude Mythos has been shared through its internal Project Glasswing on the grounds that the model carries high risk, OpenAI is moving ahead with a broader product structure aimed at real-world enterprise deployment.
Two cyber-focused models and more than 20 partners
OpenAI said Daybreak is not a single-model product. It combines OpenAI models, Codex, and a network of cybersecurity partners. The package includes GPT-5.5 with Trusted Access for Cyber and GPT-5.5-Cyber, with the latter beginning a gradual rollout last week.
The confirmed partner list spans security infrastructure and software supply chains, including Cloudflare, Cisco, CrowdStrike, Palo Alto Networks, Oracle, Zscaler, Akamai, Fortinet, Intel, Qualys, Rapid7, Tenable, Trail of Bits, SpecterOps, SentinelOne, Okta, Netskope, Snyk, Gen Digital, Semgrep, and Socket. In total, OpenAI named more than 20 organizations. The company also said it is working with industry and government partners to prepare future deployment of models with stronger cyber operations capabilities.
According to OpenAI, GPT-5.5-Cyber is meant for professional workflows in authorized settings, including approved red-team exercises, penetration testing, and controlled validation environments. OpenAI added that the model’s performance on cyberattack benchmark tests is broadly comparable to Anthropic’s Mythos.
Codex Security centers on four layers
The technical core of Daybreak is a four-part capability stack inside the Codex Security agent. First comes repository-specific threat modeling. Instead of applying a generic template, the system analyzes a company’s actual code and builds a threat model tied to that organization’s architecture. Second is review of real attack paths. Rather than listing possible weaknesses, it simulates how an attacker could move through a system and identifies which paths appear realistically exploitable.
The third layer validates issues inside an isolated environment so the detection process does not interfere with production systems. The fourth generates remediation suggestions, which are then left for human review before adoption. OpenAI’s stated goal is to compress work that can take security teams days—threat modeling and penetration testing���into a workflow an AI agent can automate while keeping final decisions in human hands.
Commercial rollout remains at an early stage
Daybreak enters a market where major positions are already being claimed. Based on currently available information, Apple, Microsoft, Google, and Amazon have joined Anthropic’s Glasswing program. Those companies also sit at the center of the global cloud and device ecosystem.
OpenAI has not disclosed pricing for Daybreak and has not published a firm launch timeline. Its current approach is to build integrations across multiple points in the security toolchain by partnering with vendors in endpoint protection, network security, vulnerability management, and identity. For now, Daybreak remains in an early rollout phase, and broader commercial adoption will depend on how those deployments develop.