OpenAI gives approved defenders an AI model built for high-risk cyber work

GPT-5.6-Cyber answered 95% of advanced requests in an internal test, but access is restricted and its benchmark performance does not consistently beat GPT-5.6 Sol

OpenAI is making GPT-5.6-Cyber available to approved defenders through the restricted Daybreak Red access tier.

OpenAI is giving approved cybersecurity specialists access to a new AI model designed to respond to advanced requests, including exploit-chain development and authentication bypass, that its general-purpose models usually reject.

GPT-5.6-Cyber is available through Daybreak Red, the more restricted of two access tiers introduced as part of an expansion of OpenAI Daybreak. The program is intended for individuals and organizations conducting authorized defensive security work.

The distinction between the tiers is significant. Daybreak Blue provides access to frontier general-purpose models, including GPT-5.6 Sol, for work such as finding vulnerabilities, reviewing code, analyzing malware, responding to incidents and validating patches.

Blue removes the standard production screening layer that can block legitimate cybersecurity requests. However, GPT-5.6 Sol can still refuse instructions involving highly dual-use activities, including testing production systems.

Daybreak Red goes further. It provides approved defenders with purpose-trained cybersecurity models for vulnerability research, exploit validation and security testing. GPT-5.6-Cyber, which is built on GPT-5.6 Sol, has been trained both to improve its performance on specialized security work and to reject fewer high-risk requests.

A model designed to answer more advanced requests

In OpenAI’s internal Advanced Cybersecurity Completion Rate evaluation, GPT-5.6-Cyber responded to 95% of requests involving scenarios such as exploit-chain development, authentication bypass and privilege escalation.

GPT-5.6 Sol completed 1.5% of the same requests with its standard safeguards enabled and 2% through Daybreak Blue. OpenAI’s earlier GPT-5.5-Cyber model completed 57.3%.

That evaluation measures whether a model responds to a request. It does not establish that the response is accurate or produces a working result.

The wider benchmark picture is mixed. OpenAI says GPT-5.6-Cyber outperformed GPT-5.6 Sol and GPT-5.5-Cyber on ExploitGym, which tests whether an agent can turn a known vulnerability into a working exploit in a controlled environment. It also beat GPT-5.6 Sol in an internal evaluation covering the discovery and assessment of previously unknown vulnerabilities.

GPT-5.6-Cyber did not lead every test. It scored below GPT-5.6 Sol in OpenAI’s Vulnerability Discovery and Report Writing evaluation, which the company attributes to the specialist model sometimes producing shorter and less detailed reports.

GPT-5.6 Sol also performed best on ExploitBench when agents were limited to the standard 300 turns. The gap narrowed when the allowance was increased to 600 turns.

OpenAI gave selected customers early access to the model. Jared Atkinson, CTO at SpecterOps, says it “has completed work in under a day that earlier models had not resolved after weeks of intermittent effort.” The statement does not identify the project involved or provide an independent measure of the result.

OpenAI reports previously unknown V8 vulnerabilities

OpenAI researchers say they used GPT-5.6-Cyber to identify two previously unknown vulnerabilities in V8, the JavaScript engine used by Chrome. The vulnerabilities could be combined to corrupt memory and escape the V8 heap sandbox.

Google fixed the first vulnerability and assigned it CVE-2026-15903. The second remains under coordinated disclosure.

The fixed vulnerability involved V8’s optimizing compiler incorrectly skipping a safety check when converting values into integers. Under certain circumstances, this could allow an attacker to read or overwrite memory belonging to other objects and potentially execute code inside Chrome’s sandbox.

OpenAI also says the model found at least five vulnerabilities in a mobile operating system, three critical vulnerabilities in a database and more than 400 issues that could allow privilege escalation in an operating system kernel. The affected products have not been named while disclosure and remediation work continues.

The broader Daybreak program includes Patch the Planet, an initiative built with Trail of Bits that combines AI-assisted research with expert review and patch development for open-source software.

Figures published by OpenAI show that 41 open-source projects are under review. The company says 858 findings have been surfaced for validation, 263 patches have been produced and tested, and 143 have been accepted by maintainers. OpenAI has committed $17 million in API credits and direct support to open-source security and the wider maintainer ecosystem.

Lower refusals come with tighter access controls

OpenAI assessed both GPT-5.6 Sol and GPT-5.6-Cyber as reaching the High threshold for cybersecurity capability under its Preparedness Framework. Neither model was assessed as reaching the Critical threshold.

The company says GPT-5.6-Cyber was not involved in the exploitation of Hugging Face described in its previous security incident update. It also states that no other model planned for release was involved. A system card containing further evaluations of GPT-5.6-Cyber will be published later.

Daybreak access is limited to approved individuals and organizations conducting authorized work. OpenAI says it uses identity verification, account security requirements, monitoring, restrictions on approved uses and legal attestations to control access.

Individual Daybreak accounts will be required to use hardware security keys from September 1, 2026. OpenAI is also encouraging customers using Codex to adopt auto-review mode, which assesses actions requiring elevated permissions before they run and can block requests considered likely to cause destructive behavior.

The company recommends Daybreak Blue for most defensive teams. Those conducting authorized vulnerability research, exploit development or red teaming can apply for Daybreak Red through the Daybreak partner program.