Daybreak Red’s 95% Is a Refusal Metric — CVE-2026-15903 Is the Check
OpenAI split Daybreak into Blue and Red and launched GPT-5.6-Cyber on Sol, claiming 95% completion on an internal advanced dual-use cyber eval versus 1.5% for Sol. The durable independent signal is a real Chrome V8 CVE chain (CVE-2026-15903) — not the refusal board alone. Trusted-access cyber, not a ChatGPT default.
Frontier labs are racing to arm defenders without flooding the open internet with Critical-class cyber agents. Daybreak Red’s 95% “completion” figure is a refusal metric, not a guarantee of exploit quality — the Chrome CVE is what makes the post more than a blog bench.
OpenAI published “Expanding Daybreak as the Cyber Defense Window Narrows,” splitting trusted cyber access into Daybreak Blue and Daybreak Red and launching GPT-5.6-Cyber — a cybersecurity-specialized model built on GPT-5.6 Sol. Blue unlocks Sol with defensive-security-calibrated safeguards; Red adds purpose-trained cyber models for authorized vulnerability research and exploit validation. On an internal Advanced Cybersecurity Completion Rate eval, GPT-5.6-Cyber completes 95.0% of advanced dual-use cyber requests versus 1.5% for Sol and 57.3% for GPT-5.5-Cyber. OpenAI says the model helped find two previously unknown V8 issues chained toward heap-sandbox escape, disclosed to Google and fixed as CVE-2026-15903.
What shipped
| Item | Detail (OpenAI Security primary) |
|---|---|
| Daybreak Blue | Frontier general models (GPT-5.6 Sol) with safeguards tailored for authorized defensive work |
| Daybreak Red | Purpose-trained cyber models for authorized exploit validation and advanced security testing |
| GPT-5.6-Cyber | Built on Sol; trained to reduce refusals and improve exploit-chain / zero-day research workflows |
| Refusal metric | 95.0% completion on Advanced Cybersecurity Completion Rate vs 1.5% Sol, 2.0% Sol+Blue, 57.3% GPT-5.5-Cyber |
| ExploitGym | GPT-5.6-Cyber outperforms Sol and GPT-5.5-Cyber on turning known vulns into working exploits (internal hardened setup) |
| Real findings | CVE-2026-15903 (Chrome V8 high-severity); additional claimed finds: ≥5 mobile OS issues, 3 critical database vulns, 400+ kernel privilege-escalation issues (disclosure ongoing) |
| Preparedness | Assessed High cyber capability, below Critical (same band as Sol; not Astra Critical territory) |
| Access | Approved individuals and orgs only; hardware security keys required for individual Daybreak accounts from September 1, 2026 |
OpenAI states GPT-5.6-Cyber was not involved in the earlier Hugging Face evaluation incident and is not an unrestricted public cyber free-for-all. Distinct from Astra Critical cyber pause (Aug 7) and GPT-5.6 Sol / Luna free-tier (Aug 6). Daybreak is trusted-access cyber distribution under identity gates — Bedrock path shipped Aug 11.
Limits
- 95% is completion/refusal, not independent exploit-quality scoring.
- Additional mobile/database/kernel find counts are lab-claimed with disclosure ongoing.
- System card and Blue vs Red enrollment scale not fully public here.
Sources
- OpenAI: Expanding Daybreak as the Cyber Defense Window Narrows (August 10, 2026)
- OpenAI Daybreak program
- CNBC: OpenAI expands Daybreak cybersecurity initiative (August 10, 2026)
- Times of AI:
openai-astra-critical-cyber-capabilities-pause(August 7)