OpenAI counters Anthropic's Mythos with Daybreak and three GPT-5.5 cyber models

OpenAI has launched Daybreak, a cybersecurity initiative arriving roughly a month after Anthropic's Project Glasswing—setting up a direct contest between the two labs over AI-powered cyber defence. Daybreak pairs OpenAI's Codex Security agent with three GPT-5.5 variants to help defenders find vulnerabilities, generate patches and verify fixes inside live repositories.

Access is tiered. Standard GPT-5.5 covers general developer work. GPT-5.5 with Trusted Access for Cyber handles secure code review, vulnerability triage, malware analysis and patch validation. GPT-5.5-Cyber sits at the top, gated behind stronger verification and reserved for authorised red teaming and penetration testing.

How Codex Security fits into the workflow

Codex Security, which rolled out in March, anchors the system. It scans a repository, builds an editable threat model, and narrows analysis to attack paths OpenAI claims are most realistic. The agent generates and tests patches in-repo, then ships evidence back to the client's tracking system. Whether it delivers on the promise of cutting hours of analysis to minutes will depend on independent testing, which has not yet surfaced.

The partner roster includes Cloudflare, Cisco, CrowdStrike, Palo Alto Networks, Oracle, Akamai, Zscaler and Fortinet. Palo Alto plans to integrate Daybreak into its Frontier AI Defense product. CrowdStrike is wiring it into Charlotte AI AgentWorks. Cisco's Anthony Grieco is backing both Daybreak and Glasswing—a notable hedge across rival initiatives.

Where OpenAI’s Daybreak and Anthropic’s Glasswing are different

The two efforts take different routes. Anthropic announced Claude Mythos a month ago and declined to release it publicly, citing offensive risk. It claimed the model surfaced thousands of zero-days, including a 27-year-old OpenBSD flaw and a 16-year-old FFmpeg bug, with Mozilla using it to patch 271 Firefox vulnerabilities. Glasswing also drew scrutiny after unauthorised parties reportedly accessed the model.

OpenAI frames Daybreak around resilience-by-design rather than continuous patching. Anthropic emphasises scale and disclosure through industry partners. Both pitches remain largely unproven outside their own announcements.