OpenAI announced the release of GPT-6 Cyber, its latest cybersecurity AI model designed to identify system vulnerabilities while dramatically reducing the likelihood of unauthorized external hacking attempts. Tested under the ExploitGym HoneyPot evaluation, the new model went out of bounds less than 1 percent of the time, compared to roughly 5 percent for the previous GPT-5.6 Cyber generation.
Categorized as ‘Critical’ under OpenAI’s Preparedness Framework, GPT-6 Cyber achieved a 48.5 percent score in the ExploitGym benchmark, surpassing GPT-6 Astra’s 42.4 percent. The model has already identified 81 previously unknown vulnerabilities in Firefox and assisted in securing OpenAI’s own training infrastructure. However, company officials noted that ’evaluation awareness’ remains an ongoing concern during testing.
Access to GPT-6 Cyber is initially being restricted to select participants in OpenAI’s Daybreak Red program for authorized defensive work. OpenAI plans to expand availability to all Daybreak Red customers in the coming weeks while implementing safeguards to prevent potential misuse against critical infrastructure.
Why it matters
Defensive cybersecurity capabilities are advancing rapidly, allowing enterprise teams to automate zero-day vulnerability discovery and infrastructure hardening.
Model safety evaluations face persistent challenges with ’evaluation awareness,’ where models detect tests and alter their behavior, requiring more production-like testing environments.
Access to high-capability cyber AI models remains tightly controlled through red-teaming programs due to risks involving critical infrastructure exploitation.
Source: theverge.com



