OpenAI expands Daybreak with GPT-5.6-Cyber

OpenAI is expanding its Daybreak program with two access tiers and introducing GPT-5.6-Cyber, a model specialized in vulnerability research, exploit development, and other authorized cybersecurity tasks.

The decision responds to a specific concern: attackers will be able to use artificial intelligence to investigate and carry out attacks faster and, in some cases, autonomously. OpenAI argues that defenders have an increasingly narrow window to prepare.

Two tiers for different defensive tasks

Daybreak is intended for approved individuals and organizations carrying out authorized security work. The program is divided into two options:

• Daybreak Blue provides access to advanced general-purpose models, including GPT-5.6 Sol, with safeguards adapted for defensive work. OpenAI recommends it as the starting point for most teams.
• Daybreak Red adds models trained specifically for vulnerability research, exploit validation, and more advanced security testing.

In practice, Blue can help review code, analyze malware, investigate incidents, detect vulnerabilities, and check whether a patch works. Red is intended for situations where a researcher needs to go further, such as building a proof of concept to demonstrate that a vulnerability allows control of a system.

GPT-5.6-Cyber reduces refusals on high-risk tasks

General-purpose models often reject cybersecurity-related requests when they could be used either to defend or to attack. That precaution can also block legitimate work, such as testing a vulnerability in a controlled environment.

OpenAI says that GPT-5.6-Cyber, available through Daybreak Red, was trained to respond better to specialized tasks and reduce those refusals in certain dual-use cases. This includes researching zero-day vulnerabilities, developing exploit chains, and validating security flaws.

In an internal OpenAI evaluation, the model completed 95.0% of the advanced requests analyzed. GPT-5.6 Sol completed 1.5% with its usual safeguards and 2.0% using Daybreak Blue. The previous model, GPT-5.5-Cyber, reached 57.3%.

These figures measure how many requests receive a response, not how many responses are correct or how many attacks would work on real systems. OpenAI also says that GPT-5.6-Cyber uses more tokens, meaning more reasoning capacity, than GPT-5.6 Sol in these tests.

It has already found flaws in real software

OpenAI says it used GPT-5.6-Cyber to investigate V8, the JavaScript engine used by Chrome. The work identified two previously unknown vulnerabilities that could be chained to corrupt memory and escape the engine's isolation. The company reported the findings to Google, and one of them was fixed under the identification CVE-2026-15903.

The flaw was in V8's optimizing compiler. An omitted security check could turn an unexpected value into a number that was too large, which could then cause the program to access memory locations outside the intended boundaries. Under certain conditions, this could allow code execution inside Chrome's isolated environment.

According to OpenAI, the model also helped locate:

• At least five vulnerabilities in a mobile operating system, including a chain that could lead from an untrusted app to privilege escalation.
• Three critical flaws in a database, one of them with a remote path to code execution.
• More than 400 vulnerabilities related to privilege escalation in an operating system kernel.

The company says it is working with its Daybreak partners and the open-source community to disclose and fix these issues. These are results reported by OpenAI, not an independent audit of all the findings.

This is not uncontrolled access

Access to Daybreak requires identity verification, account protection, monitoring, usage restrictions, and legal commitments covering authorized work. OpenAI also recommends running these models in isolated environments, without unnecessary connections to production systems or the internet.

The announced measures include preferential use of auto-review in Codex, a feature that reviews actions with elevated permissions before executing them, and a requirement to use physical security keys for individual Daybreak accounts starting September 1, 2026.

OpenAI also recommends defining in advance which systems the agent can access, limiting its permissions, and maintaining human oversight in the highest-risk workflows. Organizations can request access to Daybreak Red if their work includes advanced research, exploit development, or authorized red teaming.

The important change is not only that there is now a model capable of finding more flaws. It is that AI models are beginning to be separated according to the risk of the tasks they can perform. For defenders, this could accelerate vulnerability detection and remediation. For providers, it also increases the responsibility to verify who has access, what the agent can do, and how its actions are stopped before they reach real systems.