“For cyber defense, it means seeing risk earlier, acting sooner, and helping make software resilient by design,” a company blog post reads.
OpenAI's Daybreak initiative
OpenAI has launched Daybreak, a cybersecurity platform that combines its large language models with the Codex agentic framework to assist organizations in identifying, patching, and validating software vulnerabilities across the development lifecycle. The company presented Daybreak as a multi-tiered offering intended to balance capability with safeguards and account-level oversight. OpenAI did not respond to CyberScoop’s request for further comment.
GPT-5.5 tiers: standard, Trusted Access for Cyber, and GPT-5.5‑Cyber
Daybreak is structured around three model tiers built on GPT-5.5. The standard GPT-5.5 model is available for general enterprise and developer tasks. A mid-tier, “GPT-5.5 with Trusted Access for Cyber,” targets defensive security workflows such as vulnerability triage, malware analysis, detection engineering, and patch validation. The highest-capability tier, GPT-5.5‑Cyber, is a more permissive variant intended for specialized uses like authorized red‑teaming and penetration testing; it is currently in preview and reserved for controlled conditions. OpenAI described the program’s approach as pairing expanded capability with “trust, verification, proportional safeguards, and accountability.”
Trusted Access for Cyber and industry partners
OpenAI has positioned Trusted Access for Cyber as a response to the dual-use nature of advanced models: the same capabilities that let defenders map relationships across codebases and accelerate fixes could be misused, the company acknowledged. Rather than centrally deciding who may defend themselves, OpenAI said it aims to enable “as many legitimate defenders as possible, with access grounded in verification, trust signals, and accountability.”
Several major technology and cybersecurity firms are already participating within the Trusted Access for Cyber framework, including Cisco, Oracle, CrowdStrike, Palo Alto Networks, Cloudflare, Fortinet, Akamai, and Zscaler. Cisco’s chief security and trust officer, Anthony Grieco, described models like GPT-5.5 as a “force multiplier for defenders,” and emphasized that the value lies in the broader enterprise framework built around the models.
Anthropic’s Project Glasswing and Claude Mythos — a competitive dynamic
Daybreak arrives weeks after Anthropic unveiled Project Glasswing, which is built around Claude Mythos Preview — a cybersecurity-focused AI system Anthropic says can autonomously identify vulnerabilities at scale. Anthropic has kept access to Mythos tightly restricted for safety and national security reasons and has not made the model commercially available. Cybersecurity experts in the United States and United Kingdom, according to CyberScoop, have described Claude Mythos as a meaningful improvement over previous frontier models in finding vulnerabilities, although debate continues about its practical impact on information security.
OpenAI publicly announced the Trusted Access for Cyber program before Anthropic’s Glasswing rollout and has since expanded its program to thousands of individuals and organizations. In April, OpenAI released GPT-5.4 Cyber, a model variant fine-tuned for cybersecurity tasks including testing and vulnerability research, and governed by Know-Your-Customer and identity verification requirements. OpenAI says it intends to make iterative improvements and to expand access to Daybreak’s most capable models over time while working with industry and government partners.
What this means for security teams, policymakers, and vendors
- Security teams: AI will accelerate portions of both defensive and offensive security tasks, but practitioners warn that faster discovery of vulnerabilities is only one piece of risk reduction. Doug Merritt, chairman & CEO of Aviatrix, said the determining factor for breach impact is how far a compromised identity can move before detection — a problem of containment that tools focused on patching do not solve.
- Policymakers and federal IT leaders: Federal CIO Greg Barbaccia told CyberScoop he sees the potential of Claude Mythos to strengthen federal cyber defenses while noting “significant uncertainties” about how such models would perform in real-world conditions. OpenAI has indicated it is cautious about exercising too much centralized control over which sectors participate in its program.
- Vendors and integrators: Jared Atkinson, CTO of SpecterOps, said defenders must continue to focus on attack-path visibility and identity exposure even as tools that find vulnerabilities mature, underscoring the need to fold model outputs into broader containment and identity management practices.
OpenAI has framed Daybreak as an iterative, partner-based deployment: the company plans to expand access to “increasingly more cyber-capable models” alongside industry and government collaborators, while calibrating identity verification and account-level controls to the risk each tier presents. The competitive contrast with Anthropic — one company expanding access under verified frameworks, the other keeping a capable model tightly restricted — leaves the field divided over how to advance capability without widening the risk surface. How operators translate model-driven discovery into hardened, identity-resilient systems will determine whether these new tools reduce breach impact or simply shift the speed of both attack and defense.
