OpenAI launches GPT-5.6-Cyber model with reduced safeguards for cybersecurity work
OpenAI launched GPT-5.6-Cyber, a cybersecurity model with significantly reduced safeguards that completes 95 percent of advanced hacking and exploitation tasks, and expanded its Daybreak program to provide approved defenders access through two tiers while partnering with major cybersecurity firms. The company acknowledged the dual-use risks of providing frontier AI models with reduced safeguards but argues democratizing such access is crucial for accelerating cyber defense.
What happened
OpenAI launched GPT-5.6-Cyber, a new model variant designed specifically for cybersecurity work with reduced safeguards compared to standard versions. The model was trained to reduce unnecessary refusals when legitimate security researchers request advanced cybersecurity tasks.
OpenAI expanded its Daybreak program into two tiers: Daybreak Blue offering access to GPT-5.6-Sol with fewer restrictions, and Daybreak Red providing access to GPT-5.6-Cyber for more advanced red-teaming work. The company also partnered with 16 major cybersecurity providers including Accenture, IBM, CrowdStrike, Cisco and Palo Alto Networks to integrate the models into their security products and services.
Reported impact
GPT-5.6-Cyber completed 95 percent of requests for exploit-chain development, authentication bypass and privilege escalation tasks, compared to 1.5 percent for standard GPT-5.6-Sol and 57.3 percent for the previous GPT-5.5-Cyber model. OpenAI acknowledged that models running with reduced safeguards carry risks beyond standard model usage, whether from misuse or misalignment.
Index analysis
The model's significantly increased capability to perform advanced hacking and exploitation tasks raises concerns about dual-use risks. OpenAI's approach relies on identity checks, account security measures, monitoring, usage restrictions and legal confirmations for approved participants, but acknowledged the inherent risks of democratizing access to frontier intelligence with reduced safeguards.
Governance controls
Access to GPT-5.6-Cyber is restricted to approved individuals and organizations through the Daybreak Red program. The program includes identity verification, account security measures, usage monitoring, usage restrictions and legal confirmations for participants conducting authorized work.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.