OpenAI released GPT-6 Cyber, a cyber-specialized variant of GPT-6 Astra. It targets authorized security work like vulnerability discovery and exploit development. The model completed 100% of general security and patching tasks in evaluations. It finished 96% of exploit development tasks and 97% of red-teaming tasks. Access is restricted via Daybreak Red with identity checks and security controls. It requires single sign-on, multi-factor authentication, and API IP allowlisting. New safeguards block high-risk activity against critical systems. It maintains safety performance outside cybersecurity, including biology refusals. It matches GPT-6 Astra on most alignment tests but shows slight differences in warning responses. It exploited a honeypot in 0.07% of cases without deployment safety instructions. Deployment safety instructions reduced honeypot exploitation to 0%.
What changed
GPT-5.6 Cyber had lower completion rates and higher out-of-scope exploitation.
What it unlocks
Performing authorized cybersecurity work like exploit development and red-teaming with fewer refusals.
- 100% general security assistance and vulnerability patching tasks
- 96% exploit development tasks
- 97% red-teaming tasks
- 99.87% safe responses on severe-risk biology requests
What you need to act on it
- Daybreak Red access with identity verification
- use restrictions and organizational security requirements
- single sign-on, multi-factor authentication, and API IP allowlisting
Sources