How much work can an AI model actually do in cybersecurity now?
OpenAI seems to answer that question, with its latest cybersecurity model having found flaws in one of the world’s most widely used browsers, Google Chrome.
OpenAI’s GPT-5.6-Cyber has uncovered two previously unknown vulnerabilities in Chrome’s V8 JavaScript engine that could be chained to escape a security sandbox. This discovery gives a clearer picture of how far AI-assisted vulnerability research has advanced, while also showing why the same capabilities can create serious security risks when placed in the wrong hands.
GPT-5.6-Cyber Found the Flaws in Chrome’s Engine
OpenAI introduced GPT-5.6-Cyber as a model specifically trained for advanced cybersecurity work. Available through the company’s Daybreak Red programme for approved defenders, it carries out tasks such as vulnerability research, exploit validation, and security testing.
After training was completed, OpenAI used the model to study selected software projects, including V8, the JavaScript engine that powers Chrome. The model then identified two vulnerabilities that had not previously been known and that could be combined to corrupt memory and escape V8’s heap sandbox.
After validating the findings and reporting them to Google through its coordinated vulnerability disclosure process, the search giant fixed one of the vulnerabilities, assigned it CVE-2026-15903, and rated it high-severity.
OpenAI said a flaw in V8’s optimizing compiler caused it to skip a safety check when converting values to integers. Under the right conditions, this could cause an unexpectedly large value to be treated as an array index, allowing an attacker to read or overwrite memory belonging to other objects.
The Model is Finding More than Browser Bugs
The Chrome research was only one part of GPT-5.6-Cyber’s work.
OpenAI said the model also identified at least five vulnerabilities in a popular mobile operating system, including a chain that could move from an untrusted application to local privilege escalation.
It found three critical vulnerabilities in a popular database, including one path to remote code execution, as well as more than 400 vulnerabilities that could lead to privilege escalation in a popular operating system kernel.
The Same Capability Can Support Attackers
GPT-5.6-Cyber was built for defensive security work, but OpenAI has deliberately trained it to handle some advanced tasks that ordinary models may refuse. On an internal evaluation covering requests such as exploit-chain development, authentication bypass and privilege escalation, GPT-5.6-Cyber completed 95% of requests, compared with 1.5% for GPT-5.6 Sol with its standard safeguards.
This level of capability is why OpenAI has restricted access to approved users and organizations and uses identity verification, monitoring, and other controls.
The ChatGPT-maker has also recently warned that its upcoming Astra model may have reached the company’s “Critical” cyber capability threshold, including the ability to discover and exploit zero-day vulnerabilities. OpenAI has since slowed parts of its model development while strengthening monitoring, alignment and containment measures.
Altogether, the Chrome findings show what specialised AI security tools can already do with real software. A model designed to help defenders find weaknesses is now capable of discovering vulnerabilities that had escaped human detection.
