OpenAI has released GPT-5.6-Cyber: willing to respond to 95% of advanced cyberattack and defense inquiries

According to Watchful AI monitoring, OpenAI has released GPT-5.6-Cyber, a version of GPT-5.6 Sol specifically designed for cybersecurity training. It enhances vulnerability discovery, exploitation, and attack chain development, while significantly reducing model refusal when faced with high-risk cybersecurity tasks.

OpenAI tested advanced tasks such as attack chain development, authentication bypass, privilege escalation, among others. GPT-5.6-Cyber responded to 95.0% of these tasks, while the previous generation GPT-5.5-Cyber was at 57.3%, and the regular GPT-5.6 Sol only at 1.5%. The test measures the willingness of the model to continue the tasks, not the success rate of the attacks.

It has also successfully discovered zero-day vulnerabilities. OpenAI used GPT-5.6-Cyber to inspect Chrome's V8 engine and found two previously unknown vulnerabilities that could be chained together to compromise memory and escape the V8 heap sandbox. One of them has been patched by Google and assigned CVE-2026-15903. OpenAI also stated that it found over 400 privilege escalation vulnerabilities in a mainstream operating system kernel.

However, this model will not be directly available to everyone. GPT-5.6-Cyber is only accessible through Daybreak Red to vetted security researchers and organizations. OpenAI rates it as having "High" network capabilities, not yet reaching the more restricted "Critical" level that would trigger stricter limitations.