OpenAI has expanded its Daybreak program with the introduction of GPT-5.6-Cyber, a specialized model for cybersecurity applications. This model is accessible through Daybreak Red, one of two new access tiers within the Daybreak initiative. The program aims to equip trusted defenders with advanced AI capabilities to address the evolving landscape of cyber threats.
GPT-5.6-Cyber is built on GPT-5.6 Sol and has been trained to enhance performance on specific cybersecurity tasks, such as identifying zero-day vulnerabilities and developing exploit chains. It also features a reduced refusal rate for certain higher-risk, dual-use cyber activities. OpenAI states that threat actors are increasingly using AI for cyberattacks, necessitating advanced tools for defenders.
Key Points
- GPT-5.6-Cyber is a cybersecurity-specific model available via Daybreak Red access.
- It is designed for authorized vulnerability research, exploit validation, and security testing.
- GPT-5.6-Cyber is trained to improve capabilities in finding zero-day vulnerabilities and developing exploit chains.
- The model reduces refusals for certain higher-risk, dual-use cyber tasks.
- An internal evaluation, Advanced Cybersecurity Completion Rate, shows GPT-5.6-Cyber completes 95.0% of requests involving exploit-chain development and other advanced scenarios.
- This compares to 1.5% for GPT-5.6 Sol and 2.0% for GPT-5.6 Sol with Daybreak Blue access.
- GPT-5.6-Cyber also outperforms GPT-5.5-Cyber, which completed 57.3% of requests.
- On ExploitGym, GPT-5.6-Cyber outperforms GPT-5.6 Sol and GPT-5.5 Cyber in turning known vulnerabilities into working exploits.
- GPT-5.6-Cyber (Daybreak Red) outperformed GPT-5.6 Sol (Daybreak Blue) on an internal evaluation for generating proof-of-concept exploits for novel zero-day vulnerabilities.
- GPT-5.6 Sol performs better than GPT-5.6-Cyber on the Vulnerability Discovery and Report Writing evaluation, which OpenAI attributes to GPT-5.6-Cyber sometimes producing shorter reports.
- On ExploitBench, GPT-5.6 Sol (Daybreak Blue) performs best in the standard 300-turn setting for developing V8 exploits.
Context
According to OpenAI, the cybersecurity landscape is changing rapidly, with threat actors increasingly leveraging AI for attacks. The Daybreak program is OpenAI's response, aiming to provide frontier intelligence to trusted defenders. The program offers two access tiers: Daybreak Blue and Daybreak Red. Daybreak Blue removes system-level safeguards present in GPT-5.6 Sol to facilitate legitimate defensive work like incident detection and vulnerability management. Daybreak Red provides access to GPT-5.6-Cyber, which is specifically trained to further reduce refusals and improve performance on highly dual-use cybersecurity prompts that even GPT-5.6 Sol might refuse.
Why It Matters
This development indicates OpenAI's strategy to address the dual-use nature of advanced AI capabilities by providing specialized models and access tiers for trusted cybersecurity professionals. For builders and researchers, it highlights the potential for AI to accelerate defensive workflows, while also underscoring the ongoing challenge of balancing capability with safety in high-stakes domains.
What To Do
- Note the specific performance differences between GPT-5.6-Cyber, GPT-5.6 Sol, and GPT-5.5-Cyber across various cybersecurity evaluations.
- Consider the implications of reduced refusal rates for advanced cybersecurity tasks in authorized defensive contexts.
- Compare the access tiers, Daybreak Blue and Daybreak Red, to understand their respective guardrail configurations and target use cases.
- Watch for further details on the Daybreak program's expansion and access criteria for trusted defenders.