OpenAI has introduced a powerful new artificial intelligence model called GPT-6 Cyber, designed to identify software vulnerabilities, test security weaknesses and help researchers prevent cyberattacks. However, its advanced hacking capabilities have raised an important question: Can AI become more powerful without becoming more dangerous?
Announced on November 5, 2026, the cybersecurity model reportedly performs significantly better than its predecessors while demonstrating improved compliance with authorized security boundaries.
OpenAI Deployment Safety Hub+1
According to OpenAI, GPT-6 Cyber achieved a 48.5% success rate on ExploitGym, a cybersecurity benchmark used to evaluate vulnerability exploitation. Its predecessor, GPT-5.6 Cyber, scored 32.1%, while GPT-6 Astra achieved 42.4%.
The company also reported major improvements in preventing unauthorized activity. During controlled testing, GPT-6 Cyber crossed security boundaries in just 0.07% of cases, compared with 5.51% for GPT-5.6 Cyber. With additional deployment safeguards enabled, no unauthorized exploitation was observed in those tests.
The model has already helped discover 81 previously unknown high-severity vulnerabilities in Mozilla Firefox. OpenAI shared its findings with Mozilla, which has released fixes for many identified weaknesses.
Researchers also used GPT-6 Cyber to reproduce 1,724 Linux kernel vulnerabilities and develop working exploits for 948 of them. The findings demonstrate how advanced AI could accelerate security research.
Despite these achievements, OpenAI has classified the model as having Critical cybersecurity capabilities, highlighting its potential risks if misused.
For now, access remains restricted to approved security researchers and organizations through OpenAI’s Daybreak Red programme.
The development highlights a growing challenge for the technology industry: strengthening digital security while preventing increasingly capable AI systems from enabling cybercrime.
