OpenAI Daybreak and GPT-5.6-Cyber Release

OpenAI has expanded the Daybreak program and introduced GPT-5.6-Cyber, a specialized model designed to provide trusted defenders with advanced cybersecurity capabilities and reduced refusals for authorized security research. This initiative aims to equip security professionals with frontier intelligence to counter the increasing speed and scale of AI-driven cyberattacks.

Daybreak Access Tiers

OpenAI has established two distinct access tiers for approved defenders to ensure they have the appropriate tools for their specific security workflows:

Daybreak Blue: This tier provides access to frontier general-purpose models, including GPT-5.6 Sol, with safeguards tailored for authorized defensive security work. It is the recommended starting point for most defenders and supports tasks such as vulnerability discovery, secure code review, malware analysis, incident response, and patch validation.
Daybreak Red: This tier provides access to purpose-trained cybersecurity models, including GPT-5.6-Cyber, for authorized vulnerability research, exploit validation, and security testing.

GPT-5.6-Cyber Capabilities and Performance

GPT-5.6-Cyber is built on GPT-5.6 Sol and specifically trained to improve performance on specialized cybersecurity tasks and reduce refusals for high-risk, dual-use tasks.

Reduced Refusal Rates

To measure the effectiveness of GPT-5.6-Cyber in reducing refusals, OpenAI used an internal "Advanced Cybersecurity Completion Rate" evaluation. This metric measures how often a model responds to requests involving exploit-chain development, authentication bypass, and privilege escalation.

GPT-5.6-Cyber (Daybreak Red): 95.0% completion rate.
GPT-5.6 Sol (Daybreak Blue): 2.0% completion rate.
GPT-5.6 Sol (with safeguards enabled): 1.5% completion rate.
GPT-5.5-Cyber (Daybreak Red): 57.3% completion rate.

Technical Benchmarks

GPT-5.6-Cyber demonstrates improved performance across several cybersecurity-focused evaluations:

ExploitGym 2: GPT-5.6-Cyber outperforms both GPT-5.6 Sol and GPT-5.5-Cyber in turning known vulnerabilities into working exploits for arbitrary code execution in controlled environments.
Zero-Day Discovery: In an internal evaluation involving open-source repositories, GPT-5.6-Cyber (Daybreak Red) outperformed GPT-5.6 Sol (Daybreak Blue) in generating proof-of-concept exploits and technical write-ups for novel zero-day vulnerabilities.
Vulnerability Discovery and Report Writing: Both GPT-5.6 Sol and GPT-5.6-Cyber showed improvement over GPT-5.5-Cyber, though GPT-5.6-Cyber performed slightly worse than GPT-5.6 Sol in this specific area, which OpenAI attributes to shorter, less detailed reports.
ExploitBench 3: In a standard 300-turn limit, GPT-5.6 Sol (Daybreak Blue) performed best and was more token-efficient. However, the performance gap narrowed when the turn limit was expanded to 600 turns.

Real-World Application and Vulnerability Research

OpenAI used GPT-5.6-Cyber to identify high-severity vulnerabilities in several major software projects:

Google Chrome (V8 Engine): The model uncovered two previously unknown vulnerabilities that could be chained to corrupt memory and escape the V8 heap sandbox. This led to the discovery of CVE-2026-15903, a high-severity vulnerability where the optimizing compiler incorrectly skipped a safety check during integer conversion, potentially allowing arbitrary code execution inside the Chrome sandbox.
Mobile Operating Systems: At least five vulnerabilities were found in a popular mobile OS, including a chain leading to local privilege escalation.
Databases: Three critical vulnerabilities were identified in a popular database, including a remote path to code execution.
OS Kernels: Over 400 vulnerabilities leading to privilege escalation were discovered in a popular operating system kernel.

Preparedness and Safety Framework

Under OpenAI's Preparedness Framework, both GPT-5.6 Sol and GPT-5.6-Cyber were assessed as reaching the "High" threshold for cybersecurity capability, but remained below the "Critical" threshold.

Access Controls and Safeguards

Access to Daybreak Blue and Daybreak Red is restricted to approved individuals and organizations through identity verification, account security, monitoring, and legal attestations. To further mitigate risk, OpenAI is implementing the following measures:

Hardware Security Keys: Required for all individual accounts in Daybreak starting September 1, 2026.
Auto-Review Mode: Encouraged for Codex users to evaluate actions requiring elevated permissions before execution.
Monitoring and Alignment: Ongoing work on improved monitoring and alignment training for future releases.

Recommended Best Practices for Defenders

OpenAI recommends that defenders using these models in the following ways:

Sandbox and Isolate: Run workflows in controlled environments without access to production systems or the open internet.
Monitor Agent Actions: Use auto-review mode for tool calls outside the Codex sandbox.
Define Scope: Use scoped permission profiles to enforce authorized boundaries.