OpenAI announces 'GPT-5.6 Cyber,' an AI specifically designed for cybersecurity, and begins offering 'GPT-5.6 Sol (unsafeguarded version)' to security researchers.
OpenAI offers the security AI '
AI for Cybersecurity Teams | OpenAI | OpenAI
https://openai.com/business/solutions/cybersecurity/
Expanding Daybreak as the Cyber Defense Window Narrows | OpenAI
https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows/
We're expanding our cybersecurity initiative Daybreak and introducing GPT-5.6-Cyber, a new model for advanced, authorized cybersecurity work.
— OpenAI (@OpenAI) August 10, 2026
As the threat landscape evolves, we're frontier intelligence in the hands of trusted putting defenders before attackers can deploy… pic.twitter.com/6o3GtxCxRA
Daybreak is a service provided for cybersecurity professionals that allows them to investigate and remediate vulnerabilities using models such as GPT-5.5 Cyber. It has been announced that Daybreak will be divided into two types: 'Daybreak Blue' and 'Daybreak Red'. Daybreak Blue provides access to the 'unprotected version of GPT-5.6 Sol,' while Daybreak Red allows users to use 'GPT-5.6 Cyber,' which is based on GPT-5.6 Sol but with improved vulnerability detection and exploit construction capabilities.
The following are the results of a survey on the 'completion rate of instructions to execute cyberattacks' for 'GPT-5.6 Sol,' 'GPT-5.6 Sol (without safeguards),' 'GPT-5.5 Cyber,' and 'GPT-5.6 Cyber.' The completion rate was 1.5% for GPT-5.6 Sol and 2.0% for GPT-5.6 Sol (without safeguards), while GPT-5.6 Cyber was able to complete 95.0% of its instructions.
The following are the results of measuring the ability to build exploits based on known vulnerabilities. GPT-5.6 Sol (without safeguards) successfully outperformed GPT-5.5 Cyber. GPT-5.6 Cyber achieved an even higher success rate.
The following is a comparison of the success rates for the task 'Assess the severity of zero-day vulnerabilities and create a technical report.' GPT-5.6 Cyber recorded the highest success rate.
On the other hand, in the 'Discovery and Reporting of Known Vulnerabilities' test, GPT-5.6 Sol (without safeguards) achieved the highest score. GPT-5.6 Cyber tended to produce reports with less detail.
The following are the results of a test to develop an exploit against the V8 JavaScript engine. Under the condition of 'limiting the number of agent turns to 300,' GPT-5.6 Sol (without safeguards) efficiently solved the task and achieved the highest score. When the number of turns was increased to 600, the difference between GPT-5.6 Sol (without safeguards) and GPT-5.6 Cyber narrowed.
According to OpenAI, GPT-5.6 Cyber is already being used to detect vulnerabilities in actual software and has discovered 'two vulnerabilities that allow escape from the V8 sandbox,' 'a vulnerability in a popular mobile OS that leads to local privilege escalation,' 'three vulnerabilities related to remote code execution in widely used databases,' and 'more than 400 vulnerabilities in the kernels of widely used operating systems.'
Please note that you need to apply via the following link to participate in Daybreak Blue and Daybreak Red.
Trusted Access For Cyber | OpenAI
https://openai.com/form/enterprise-trusted-access-for-cyber/
Related Posts:
