OpenAI on Monday announced a new cybersecurity-focused AI model named GPT-5.6-Cyber, as well as the expansion of its Daybreak Cyber Partner program.
The AI giant says GPT-5.6-Cyber is designed for “advanced, authorized cybersecurity work”. Built on GPT-5.6-Sol, the new model has been trained for specialized tasks such as finding zero-days and creating exploit chains. It’s also configured with a lower refusal rate for risky dual-use tasks.
According to OpenAI, while GPT-5.6-Sol is highly advanced, its refusal rate is high for penetration testing against production systems and other dual-use cyber prompts that could be used for both good and evil.
GPT-5.6-Cyber should address these issues, and testing conducted by the vendor showed a 95% completion rate for prompts involving exploit chain development, privilege escalation, and authentication bypass, compared with 1.5% for GPT‑5.6 Sol.
“It also completes more requests than GPT‑5.5‑Cyber, which completes only 57.3% of requests, addressing feedback from security researchers who encountered persistent refusals with the earlier model,” OpenAI noted.
GPT-5.6-Cyber has superior performance in areas such as developing arbitrary code execution exploits and finding and assessing known vulnerabilities and new zero-days, according to OpenAI.
The company says the model has discovered a high-severity vulnerability in the V8 JavaScript engine used by Chrome. It has also reportedly found security flaws in an unnamed mobile operating system, database, and operating system kernel.
To prevent abuse of the new GPT-5.6-Cyber model, OpenAI will offer it only to trusted partners through an expansion of its Daybreak program, which enables organizations to build, co-sell, and deliver cybersecurity solutions with AI developed by OpenAI.
The company announced on Monday that Daybreak is being expanded with two access tiers: Daybreak Blue, which provides access to Sol and other general-purpose frontier models with guardrails customized for defensive cybersecurity work, and Daybreak Red, which provides access to cybersecurity models such as the new GPT-5.6-Cyber.
Daybreak has been continuously expanded. The list of partners includes major tech and services companies such as Accenture, Capgemini, EY, IBM, KPMG, and PwC, and cybersecurity giants such as Palo Alto Networks, Sophos, CrowdStrike, Fortinet, Akamai, and Cloudflare.
The latest announcements come just days after OpenAI flagged its upcoming AI model Astra for potentially reaching a ‘critical’ cybersecurity risk threshold.
GPT-5.6-Sol has peaked at the ‘high’ risk threshold, but Astra could reach ‘critical’, which means it could autonomously build zero-day exploits and could independently design and execute end-to-end cyberattacks based only on a high-level goal.
Recent incidents in which models made by OpenAI, Anthropic and Meta hacked real organizations during cybersecurity testing have raised serious concerns about the potential risks of autonomous AI capabilities.
Related: OpenAI Refocuses Cybersecurity Efforts on Patching Over Discovery
Related: Critical One-Click Vulnerability in Atlassian’s Rovo AI Exposed Enterprise Data
