OpenAI launches restricted model for security work it won't let general users access

OpenAI unveiled GPT-5.6 Cyber, a specialized model designed explicitly for security professionals that the company has kept behind multiple layers of gatekeeping. Unlike its general-purpose models, Cyber is trained to tackle advanced offensive security tasks—exploit development, zero-day discovery, and authentication bypass—that standard AI assistants refuse to attempt.

The model is available exclusively through Daybreak Red, the highest tier of a newly formalized two-tier gated program. In internal testing on tasks like privilege escalation and exploit chains, Cyber handled 95% of requests, compared to 1.5% under standard protections and 2% on the lower-restriction Daybreak Blue tier for its sister model, GPT-5.6 Sol.

How the two-tier system works#

The architecture splits responsibilities along security workflows. Daybreak Blue grants approved defenders access to Sol with standard safety filters lifted, suitable for secure code review, malware analysis, incident response, and vulnerability discovery. Red unlocks Cyber itself, which tackles genuinely offensive work that Sol continues to refuse.

In a test case, both models were asked to generate code for bypassing macOS Keychain prompts and decrypting Chrome cookies. Sol rejected the request across both Blue and standard modes; Cyber delivered the code through Red.

Access to either tier requires identity verification, legal attestations, and account monitoring. Starting September 1, 2026, individual users must enable hardware security keys.

Real-world discoveries#

OpenAI used Cyber internally to audit Google's V8 JavaScript engine, uncovering two previously unknown vulnerabilities that could chain together to escape the V8 heap sandbox—Google addressed them under CVE-2026-15903. The model also identified three critical flaws in a popular database with remote code execution paths, at least five in a mobile OS, and over 400 potential privilege-escalation issues in an operating-system kernel. The company has withheld names pending coordinated disclosure.

According to OpenAI's Preparedness Framework, Cyber reached "High" for cybersecurity capability but stopped short of the "Critical" threshold.

Where Cyber underperforms#

Despite its strengths on exploit development, Cyber doesn't uniformly outpace Sol. When tasked with discovering vulnerabilities and documenting findings, Sol produced more thorough, detailed reports. Cyber also consumed more tokens on benchmark tests and fell behind Sol when reasoning was capped at 300 turns—the gap narrowed only at higher token budgets, suggesting longer research sessions will cost more.