GPT-5.6-Cyber Answers 95% of Advanced Exploit Requests

A cyber model built to finish high-risk defensive work

GPT-5.6-Cyber completed 95.0% of advanced cybersecurity requests that general models usually refuse. On the same internal evaluation, GPT-5.6 Sol completed 1.5%, Sol with Daybreak Blue access reached 2.0%, and the earlier GPT-5.5-Cyber reached 57.3%. The important change is not merely faster code generation. It is the proportion of authorized, dual-use security work the model will carry through, including exploit-chain development, authentication bypass, and privilege escalation.

“We’re expanding our cybersecurity initiative Daybreak and introducing GPT-5.6-Cyber.” — OpenAI

The source tweet, posted August 10, paired the model launch with an expansion of OpenAI’s Daybreak access program. OpenAI’s account is the lab’s official channel for model releases, product changes, and safety research. The linked technical report describes GPT-5.6-Cyber as a GPT-5.6 Sol derivative trained to reduce unnecessary refusals on specialized tasks such as zero-day discovery and exploit-chain construction.

Two unknown V8 flaws move the claim beyond a benchmark

OpenAI says it used the model to find two previously unknown vulnerabilities in V8, the JavaScript engine inside Chrome. Together, the flaws could corrupt memory and escape V8’s heap sandbox. Google fixed the high-severity issue designated CVE-2026-15903 after coordinated disclosure. OpenAI also reports at least five flaws in a popular mobile operating system, three critical vulnerabilities in a popular database, and more than 400 privilege-escalation issues in a widely used operating-system kernel. The affected products have not all been named because disclosure and remediation are still under way.

The results include meaningful limits. GPT-5.6-Cyber beat Sol and GPT-5.5-Cyber on ExploitGym, which asks agents to turn known vulnerabilities into working code-execution exploits. Yet GPT-5.6 Sol remained the most token-efficient model on the harder ExploitBench setting capped at 300 turns. Sol also scored better on an internal vulnerability-discovery and report-writing evaluation because the cyber model sometimes produced shorter, less detailed reports. When the exploit task was extended to 600 turns, the gap narrowed.

Restricted access is part of the product

GPT-5.6-Cyber is not being released through an unrestricted public API. Daybreak Blue gives approved defenders GPT-5.6 Sol for broad defensive work, while Daybreak Red provides the specialized model for advanced vulnerability research, exploit validation, and security testing. OpenAI says access requires identity verification, account protections, monitoring, approved-use restrictions, and legal attestations. Individual Daybreak accounts must also adopt hardware security keys beginning September 1, 2026.

Three things now deserve scrutiny: how much evidence the promised system card provides about the internal evaluations, how quickly the unnamed mobile, database, and kernel flaws are patched, and whether Red access can produce measurable defensive gains without enabling misuse. A 95.0% completion rate is striking, but the stronger test is whether those completed requests turn into safely disclosed and repaired vulnerabilities.