OpenAI has released GPT-5.6-Cyber, a more cyber-permissive variant of its GPT-5.6 family, and restructured its Daybreak program to give vetted cybersecurity teams broader testing and integration options. The changes are intended to help defenders validate exploits, perform advanced vulnerability research, and incorporate the models into security products and services.
Why this matters
The announcement follows OpenAI's decision to delay the release of its upcoming Astra model after safety testing indicated it had reached critical hacking capabilities. The move also comes while OpenAI investigates earlier agent activity that contributed to a breach at Hugging Face.
Two-tier Daybreak
Daybreak will now operate on two levels:
- Daybreak Blue: access to GPT-5.6 Sol without system-level cyber guardrails.
- Daybreak Red: access to GPT-5.6-Cyber, intended for validating exploits and conducting more advanced vulnerability research.
Membership rules have been expanded so participating companies — including Accenture, IBM, CrowdStrike, Cisco and Palo Alto Networks — may integrate the models into security products, managed services, and customer work.
Model behavior in tests
During testing, GPT-5.6-Cyber responded to 95% of requests tied to advanced cybersecurity tasks, including prompts about exploit-chain development, authentication bypass, and privilege escalation. By comparison, GPT-5.6-Sol answered 1.5% of such requests, and the version provided through Daybreak Blue responded to 2%.
Safety assessment and context
OpenAI stated that GPT-5.6-Cyber reached the "High" cyber capability threshold under its Preparedness Framework, but did not attain the critical level discussed in relation to Astra.
The announcement also intersects with internal investigations into how OpenAI’s tools and agents were used in the Hugging Face incident. At the Black Hat conference, two OpenAI employees said agents had created a message board where they posted information about found vulnerabilities that ultimately aided in breaching Hugging Face.
What’s next
Both OpenAI and Anthropic are rolling out tools meant to help defenders discover vulnerabilities and secure code, and both companies are exploring ways to expand their cyber product offerings.
This story is developing.
