OpenAI has launched GPT-5.6-Cyber, a model designed for cybersecurity tasks with a 95% completion rate for high-risk requests, accessible only through its Daybreak program. This follows a recent pause on Astra due to critical cyber risks, highlighting OpenAI's cautious approach to model deployment.
TL;DR
- OpenAI releases GPT-5.6-Cyber with a 95% completion rate for cybersecurity tasks.
- Model available exclusively through Daybreak, a vetted cybersecurity program.
- OpenAI pauses Astra due to critical cyber risks, emphasizing controlled model deployment.
What happened
OpenAI has introduced GPT-5.6-Cyber, a model built on GPT-5.6 Sol and trained for zero-day discovery and exploit-chain development. The model is designed to refuse fewer high-risk dual-use cyber requests and is only accessible through Daybreak, OpenAI's vetted cybersecurity program. This release comes three days after OpenAI delayed Astra due to concerns over critical cyber capabilities. GPT-5.6-Cyber achieves a 95% Advanced Cybersecurity Completion Rate, significantly higher than its predecessor, GPT-5.5-Cyber, which had a 57.3% rate. The model has already identified several vulnerabilities, including two high-severity issues in Google's V8 JavaScript engine, which have been fixed and assigned CVE-2026-15903.
Why it matters
The release of GPT-5.6-Cyber provides cybersecurity professionals with a powerful tool for vulnerability research and exploit validation. This model is part of OpenAI's strategy to equip defenders with advanced AI tools to counter potential cyber threats. The high completion rate indicates improved compliance with cybersecurity tasks, addressing concerns about model refusal rates. However, the model's performance on certain benchmarks is mixed, with some tasks performed better by the general-purpose Sol model. OpenAI's cautious approach to model deployment, as seen with the pause on Astra, underscores the importance of responsible AI development and controlled access to advanced models.
Key facts
- GPT-5.6-Cyber is built on GPT-5.6 Sol and is designed for cybersecurity tasks.
- The model achieves a 95% Advanced Cybersecurity Completion Rate.
- Access to GPT-5.6-Cyber is restricted to Daybreak, a vetted cybersecurity program.
- OpenAI paused Astra due to concerns over critical cyber capabilities.
- GPT-5.6-Cyber has identified two high-severity vulnerabilities in Google's V8 JavaScript engine, fixed as CVE-2026-15903.
- The model is available to government-approved partners and select cybersecurity firms.
- Daybreak access requires identity verification, account security, monitoring, and legal attestations.
- OpenAI plans to enforce hardware security keys for Daybreak accounts starting September 1, 2026.
Context
OpenAI's release of GPT-5.6-Cyber highlights the growing importance of AI in cybersecurity. As cyber threats become more sophisticated, the need for advanced tools to identify and mitigate vulnerabilities is crucial. OpenAI's cautious approach to model deployment reflects the broader industry's focus on responsible AI development and the need to balance the benefits of advanced models with potential risks. The release of GPT-5.6-Cyber is part of a broader trend in the AI industry to equip cybersecurity professionals with powerful tools to defend against emerging threats. This development underscores the critical role of AI in enhancing cybersecurity capabilities and the ongoing efforts to ensure the responsible use of advanced models.
