OpenAI Launches GPT-5.6-Cyber, a Specialized Model That Answers 95% of Blocked Security Queries

OpenAI has released GPT-5.6-Cyber, a specialized model built to find vulnerabilities and construct exploit chains, as part of an expansion of its Daybreak cybersecurity program. The model answers 95 percent of sensitive security queries in OpenAI's internal "Advanced Cybersecurity Completion Rate" benchmark — queries that standard safety-tuned models block almost entirely.

The Daybreak program now splits into two access tiers. Daybreak Blue targets defensive work — malware analysis, vulnerability detection, incident response — and runs on GPT-5.6 Sol with tailored safeguards. Daybreak Red is aimed at offensive security researchers doing exploit validation and penetration testing, and it's the only tier that unlocks GPT-5.6-Cyber.

Benchmark gap between safety and capability

On OpenAI's internal benchmark covering exploit chain development, authentication bypass, and privilege escalation scenarios, the numbers show a sharp divide by access tier and safety configuration:

• GPT-5.6-Cyber (Daybreak Red): 95%
• GPT-5.5-Cyber (previous generation): 57.3%
• GPT-5.6 Sol with Daybreak Blue safeguards: 2%
• GPT-5.6 Sol with standard safety measures: 1.5%

In one test case, models had to develop a WebSocket authentication bypass for an internal admin panel. Only GPT-5.6-Cyber running on Daybreak Red produced working exploit code; every other configuration refused. OpenAI also reports that GPT-5.6-Cyber outperforms both GPT-5.6 Sol and GPT-5.5-Cyber on ExploitGym, a benchmark measuring how reliably a model converts known vulnerabilities into functional exploits.

Real vulnerabilities found

OpenAI says it used GPT-5.6-Cyber to analyze V8, Chrome's JavaScript engine, and identified two previously unknown vulnerabilities that can be chained to corrupt memory and bypass the V8 heap sandbox. Google patched the flaws following coordinated disclosure; they're tracked as CVE-2026-15903. OpenAI also claims the model found at least five vulnerabilities in an unnamed "popular mobile operating system," including a chain that could let an app escalate from restricted access to full administrator control of a device. Disclosure and fixes for those issues are reportedly in progress with Daybreak partners and the open-source community.

Access controls

Entry into either Daybreak tier requires identity verification, account security measures, ongoing monitoring, and legal declarations. Hardware security keys become mandatory for all Daybreak accounts starting September 1, 2026. OpenAI recommends running security workflows in isolated sandboxes and using Codex's Auto-Review mode, which flags actions requiring elevated privileges before execution.

Under OpenAI's Preparedness Framework, GPT-5.6-Cyber is rated "High" for cybersecurity capability — one tier below "Critical." OpenAI notes its upcoming Astra model is "potentially" expected to reach the Critical threshold.

What this means

The jump from 57.3 percent to 95 percent completion rate between GPT-5.5-Cyber and GPT-5.6-Cyber in a single generation is the real signal here — offensive AI capability is compounding fast, not plateauing. Gating this model behind identity verification and hardware keys is a reasonable containment step, but it doesn't change the underlying dynamic: the capability now exists in a deployable form, and access controls are a policy layer, not a technical ceiling. OpenAI's framing — give defenders a head start — assumes defenders will actually use tools like this faster and more broadly than attackers exploit leaked access, stolen credentials, or comparable open models that lack these guardrails entirely. The Preparedness Framework rating of "High" rather than "Critical" is doing a lot of work in that argument, and it's worth watching whether Astra, OpenAI's next model, crosses that line.