OpenAI is expanding its Daybreak program with two access levels for authorized defenders and introducing GPT-5.6-Cyber, a model specialized in vulnerability research, exploit validation, and security testing. The initiative arrives at a time when artificial intelligence can accelerate attacks and reduce organizations’ preparation time.
Two Levels for Cybersecurity Defense
OpenAI’s proposal aims to put advanced models in the hands of teams performing legitimate security work. The company divides access into two options:
- Daybreak Blue: offers general-purpose frontier models, including GPT-5.6 Sol, with protections adapted to authorized defensive tasks. It is intended as the starting point for most teams.
- Daybreak Red: includes models trained specifically for advanced vulnerability research, exploit validation, and authorized security testing.
Daybreak Blue can be used for activities such as secure code review, malware analysis, incident response, vulnerability management, research, and patch validation. Why separate access levels? Because not every team needs the same level of capability or faces the same risks.
GPT-5.6-Cyber Reduces Refusals in Advanced Tasks
GPT-5.6-Cyber is available through Daybreak Red and is based on GPT-5.6 Sol. OpenAI trained it to improve specialized tasks such as finding zero-day vulnerabilities, developing exploit chains, and technically assessing flaws.
The main difference is not simply that the model can solve more complex problems. It is also designed to reject fewer requests related to dual-use activities, always within a restricted-access environment for approved researchers and organizations.
According to an internal OpenAI evaluation called Advanced Cybersecurity Completion Rate, GPT-5.6-Cyber completed 95% of the advanced requests evaluated. GPT-5.6 Sol reached 1.5% with its general protections and 2% through Daybreak Blue. GPT-5.5-Cyber reached 57.3%.
These numbers need to be read in context: they are results from internal tests and, on their own, do not amount to a universal measure of safety or effectiveness. OpenAI also notes that GPT-5.6-Cyber uses a larger reasoning budget, which may increase token consumption.
Mixed Results in Security Tests
In ExploitGym, an evaluation that measures whether an agent can turn known vulnerabilities into working exploits within controlled environments, GPT-5.6-Cyber outperformed GPT-5.6 Sol and GPT-5.5-Cyber.
It also achieved better results than GPT-5.6 Sol in an internal test involving zero-day vulnerabilities. In that scenario, the models had to analyze code repositories, create proof-of-concept exploits, and explain the technical impact of their findings.
However, it did not win in every evaluation. In a vulnerability discovery and reporting test, GPT-5.6 Sol performed better. OpenAI attributes part of the difference to the fact that GPT-5.6-Cyber sometimes produces shorter, less detailed reports.
In ExploitBench, which evaluates exploit creation for V8 vulnerabilities with less information and more active protections, GPT-5.6 Sol was more efficient in the standard 300-turn scenario. When the limit increased to 600 turns, the difference between the two models narrowed.
The conclusion is not that one model is better at everything. Each system seems to excel depending on the type of research, the amount of context, and the time available.
Vulnerabilities Found in Real-World Software
OpenAI says it used GPT-5.6-Cyber to study real software projects, including V8, the JavaScript engine used by Chrome. The research identified two vulnerabilities that could be chained to corrupt memory and escape V8’s heap isolation.
The researchers validated the findings and reported them to Google through coordinated disclosure. One of the vulnerabilities received the identifier CVE-2026-15903 and was classified as high severity.
According to the description published by OpenAI, V8’s optimizing compiler skipped a security check when converting certain values into integers. This could generate an unexpectedly large array index and allow reads or writes outside the intended boundaries. Under certain conditions, the issue could contribute to code execution within Chrome’s sandboxed environment.
The company also says it found high-severity vulnerabilities in a mobile operating system, a popular database, and an operating system kernel. The results mentioned include a chain that could lead from an untrusted application to local privilege escalation, three critical flaws in a database, and more than 400 vulnerabilities related to privilege escalation in a kernel.
OpenAI says it is working with its Daybreak partners and the open-source community to disclose and fix these issues. In this context, the most important value of these systems is not producing attacks, but helping discover flaws before they are exploited.
The Model Remains Below the Critical Threshold
Within its Preparedness Framework, OpenAI classified GPT-5.6 Sol as High in cybersecurity capability, but below the Critical level. GPT-5.6-Cyber also reached the High threshold without reaching Critical.
This means the model is improving at tasks for which it was specifically trained, but the company itself does not consider it to have reached a critical level of cyber capability. OpenAI also clarifies that GPT-5.6-Cyber was not involved in exploiting the Hugging Face incident mentioned in its update and that no other models related to that case are planned for an upcoming release.
The company says it will publish a technical report or system card with additional evaluations later.
Restricted Access and Additional Controls
Daybreak Blue and Daybreak Red are available to approved individuals and organizations performing authorized work. Access includes identity verification, account security controls, monitoring, usage restrictions, and legal statements about the purpose of the activities.
OpenAI also announced several additional measures:
- It recommends using Codex’s automatic review mode instead of full-access mode. This system reviews actions requiring elevated permissions before executing them.
- It will require physical security keys for all individual Daybreak accounts starting September 1, 2026.
- It plans to add monitoring improvements over the following weeks.
- It will prioritize training and alignment testing in future Daybreak versions.
- It updated its Codex documentation with practices for keeping agents within defined security boundaries.
Usage recommendations include running workflows in isolated environments, avoiding direct access to sensitive production systems, monitoring agent actions, and clearly defining which systems and operations are authorized.
Organizations can also customize their review policies to adapt controls to their internal processes. For most defenders, OpenAI recommends starting with Daybreak Blue. Teams working on advanced vulnerability research, exploit development, or red-team exercises can request access to Daybreak Red.
The race between attackers and defenders using artificial intelligence is no longer a distant concern. The practical question is how to deliver powerful capabilities without turning them into an open door for abuse. OpenAI is betting on restricted access, controlled environments, and human oversight, but the effectiveness of that model will depend on whether its controls evolve as quickly as its technical capabilities.
Original Source
https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows