OpenAI launches GPT-5.6-Cyber via Daybreak Red — cyber defense access stays gated

OpenAI just put a sharper cyber knife in trusted hands and left everyone else staring at the rope. In today’s Daybreak expansion post, the lab introduced GPT-5.6-Cyber for authorized vulnerability research and split Daybreak into Blue and Red tiers. The claim is blunt: defenders need frontier cyber help before attackers industrialize it. Public ChatGPT is not the delivery channel.

What GPT-5.6-Cyber is trained to do

GPT-5.6-Cyber is built on GPT-5.6 Sol and fine-tuned for specialized cybersecurity work: finding zero-days, building exploit chains, and other high-friction dual-use tasks where the general model still refuses. OpenAI says it has already used the model in real research, including work that uncovered previously unknown issues in Chrome’s V8 engine. Google fixed one high-severity case as CVE-2026-15903 after coordinated disclosure. The company also cites findings across a popular mobile OS, a major database, and a widely used kernel.

Preparedness framing stays below Critical: OpenAI rates the model High for cyber capability, same band as Sol, and says it will publish a fuller system card later. Vendor evals, not an outside audit.

Daybreak Red vs Blue access

Daybreak Blue is the default on-ramp. Approved defenders get frontier general-purpose models, including Sol, with system-level cyber guardrails removed so legitimate incident response, vuln management, and assessments stop dying on policy filters.

Daybreak Red is the higher-risk lane. It unlocks purpose-trained cyber models, including GPT-5.6-Cyber, for authorized vulnerability research, exploit validation, and security testing. Blue is “use the frontier model without the nanny.” Red is “use the model trained to say yes more often on dual-use cyber.”

Refusal-rate claims vs Sol

OpenAI’s internal Advanced Cybersecurity Completion Rate is the headline number. It scores whether a model will engage on exploit-chain development, authentication bypass, privilege escalation, and similar prompts. GPT-5.6-Cyber completes 95.0% of those requests. GPT-5.6 Sol sits at 1.5%. Sol under Daybreak Blue only reaches 2.0%. Prior GPT-5.5-Cyber lands at 57.3%.

On ExploitGym, OpenAI says Cyber beats both Sol and 5.5-Cyber at turning known vulns into working exploits in controlled settings. Results are mixed elsewhere: Sol can still win on some report-writing and token-efficiency benches. The product intent is clear anyway — fewer dead ends for authorized red teams, not a free public offensive API.

Who can actually request access

Access is for approved individuals and organizations doing authorized work. Controls include identity verification, account security, monitoring, approved-use restrictions, and legal attestations. Individual Daybreak accounts must adopt hardware security keys starting September 1, 2026. OpenAI is pushing Codex users toward auto-review instead of full-access execution and says improved monitoring is coming in weeks.

Apply path: openai.com/daybreak/partners. Blue first for most defenders; Red only if your authorized scope includes advanced vuln research, exploit development, or red teaming.

Analysis: this is the same dual-use bargain Google made with Flash Cyber — sell the capability to people you already trust, and keep the refusal wall high for everyone else. The 95% completion claim is OpenAI grading OpenAI. The V8 CVE is the concrete receipt. Watch whether Red stays a narrow partner club or becomes the quiet default for every big MSSP that can pass KYC.