OpenAI unveils a new ChatGPT designed to hunt for security vulnerabilities... and exploit them

OpenAI has just unveiled GPT-5.6-Cyber, a new version of its AI built for cybersecurity. Based on GPT-5.6 Sol, this model specializes in finding zero-day vulnerabilities and building exploit chains.It is set to succeed GPT-5.5-Cyber, which was announced only last May.

The improvements to GPT-5.6-Cyber

OpenAI measures its model’s performance using a metric of its own, namely the response rate for requests involving cybersecurity-sensitive matters. In this regard, the AI performs considerably better than its predecessor, responding to 95% of requests, compared with 57.3% for the previous version. For comparison, the public version of ChatGPT only deigns to respond to 1.5 to 2% of requests. In plain terms, OpenAI has not merely made its AI more capable. Above all, it has reduced its refusals to respond to offensive requests, provided the user has been duly verified by the company. This change is in response to “feedback from cybersecurity specialists who found that the previous version was still refusing certain legitimate requests too often”, OpenAI tells us.

We’re expanding our cybersecurity initiative Daybreak and introducing GPT-5.6-Cyber, a new model for advanced, authorized cybersecurity work.

As the threat landscape evolves, we’re putting frontier intelligence in the hands of trusted defenders before attackers can deploy… pic.twitter.com/6o3GtxCxRA

— OpenAI (@OpenAI) August 10, 2026

Among GPT-5.6-Cyber’s achievements is the discovery of“a series of vulnerabilities in the V8 codebase”, the JavaScript engine used by Chrome.When exploited simultaneously, “these flaws could allow an attacker to disrupt the system’s memory and bypass certain security protections”. The V8 engine is a prime target for cybercriminals. This engine is the core that powers JavaScript on virtually every modern website. It is therefore mechanically the component in which the majority of Chrome’s discovered flaws are concentrated. Last year, Google also patched 8 actively exploited zero-day vulnerabilities in Chrome.

Also read: ChatGPT and Claude slipped beyond all control – AIs have launched 2 new “real-world” cyberattacks

A program reserved for cybersecurity experts

As expected, GPT-5.6-Cyber is not available to all internet users. It is reserved for members of the Daybreak program, a platform entirely dedicated to protecting software against cyberattacks with the help of AI. Access to the program is strictly limited to consulting firms and approved cybersecurity vendors.

“In response to the rapidly evolving threat landscape, OpenAI seeks to make its most advanced technologies available to trusted professionals before malicious actors can use AI at scale for offensive purposes,” OpenAI said.

OpenAI emphasizes that access “is restricted to pre-approved individuals and organizations operating within an authorized framework”, subject to “identity checks, account security measures, and monitoring of model usage”. OpenAI also plans “restrictions on permitted uses and legal obligations”. Starting September 1, 2026, every individual Daybreak account will even need to use a hardware security key to log in. OpenAI is determined to prevent cybercriminals from gaining access to its new AI.

Also read: OpenAI makes an unexpected revelation about the cyberattack launched by a cheating ChatGPT

A two-track Daybreak program

The Daybreak program is now divided into two distinct tracks. Daybreak Blue provides access to general-purpose models such as GPT-5.6 Sol, with lighter safeguards. This track is primarily intended for users seeking defensive tools, with standout features such as vulnerability research, secure code analysis, malware analysis, and patch validation before deployment.It is “recommended as a starting point for most professionals”, OpenAI tells us.

The second component, called Daybreak Red, features models that specialize more in developing exploitation tactics, such as GPT-5.6-Cyber. This program is designed to carry out authorized penetration tests and validate complete exploit chains. It is “intended for experienced professionals who work, in an authorized setting, on complex security testing”, the startup detailed in an email sent to 01net. Several major cybersecurity companies are currently collaborating with OpenAI under DayBreak, including Accenture, IBM, Capgemini, CrowdStrike, Cisco, Palo Alto Networks and Cloudflare.

please consider using our models to help defend your systems https://t.co/oqs3qfWmtN

— Sam Altman (@sama) August 10, 2026

This announcement comes as OpenAI has lost control of several AI models during cybersecurity testing, including experiments conducted by third parties. The startup noted that the AIs went beyond the intended scope during cybersecurity testing by accessing the internet. Once online, they attacked several companies, including an open-source platform. In response to these incidents, OpenAI chose to slow the development of its next flagship model, Astra.Even as it slows down, the king of AI continues to unveil new tools designed for cybersecurity and reserved for carefully selected entities.

👉🏻 Follow tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.