OpenAI Launches GPT-5.6-Cyber and Daybreak Tiers for Advanced Security

According to OpenAI, the model reduces refusals for high-risk “dual-use” requests that general-purpose models typically block, enabling vetted security teams to identify zero-day vulnerabilities and develop exploit chains.

The model is a fine-tuned version of GPT-5.6 Sol, which launched in June. OpenAI researcher Eric Wallace stated on X that GPT-5.6-Cyber represents the company’s first large-scale attempt at directly improving capabilities for advanced cybersecurity tasks such as exploit development.

In an internal benchmark called Advanced Cybersecurity Completion Rate, GPT-5.6-Cyber completed 95% of tasks involving privilege escalation, authentication bypass, and exploit-chain development. This outperforms its predecessor, GPT-5.5-Cyber, which had a 57.3% completion rate, and the standard GPT-5.6 Sol model, which completed only 1.5% of the same tasks when safeguards were applied, according to an OpenAI launch blog post.

Daybreak Red and Blue Access Tiers

OpenAI is not making GPT-5.6-Cyber available to all ChatGPT or API customers. Access requires acceptance into the Daybreak Red tier of the Daybreak cybersecurity program, which is reserved for approved security teams performing authorized work such as penetration testing and red-team exercises on systems they own or have permission to test.

Organizations must apply through Daybreak Access, confirming their work is lawful and defensive. OpenAI requires applicants to demonstrate a professional security program featuring single sign-on, multifactor authentication, role-based access, and a documented incident-response process. Applicants must also provide a recognized security certification, such as ISO 27001 or SOC 2 Type II.

A second, broader tier called Daybreak Blue provides vetted defenders access to general models like GPT-5.6 Sol with adjusted guardrails. OpenAI states Daybreak Blue is intended for secure-code review, malware analysis, and patch validation, rather than the specialized exploit work supported by the Red tier.

GPT-5.6-Cyber costs $12.50 per million input tokens and $75 per million output tokens, while cached input is priced at $1.25 per million tokens. These rates exceed those for GPT-5.6 Sol in the Daybreak table, where short-context use is priced at $5 per million input tokens and $30 per million output tokens.

Verified Zero-Day Discoveries

OpenAI reports that researchers used GPT-5.6-Cyber to find two previously unknown vulnerabilities in V8, the JavaScript engine used by Chrome. These flaws could be chained to escape the V8 heap sandbox and corrupt memory. Google has since fixed one of these vulnerabilities, assigned as CVE-2026-15903, which involved an out-of-bounds array index in V8’s optimizing compiler.

OpenAI further claims the model contributed to finding five vulnerabilities in a popular mobile operating system, three critical flaws in a popular database, and over 400 privilege escalation vulnerabilities in a popular operating-system kernel. The company stated these disclosures are currently being coordinated.

Jared Atkinson, CTO of SpecterOps, stated that GPT-5.6-Cyber is materially improving our specialist vulnerability-research workflows, noting that the model completed some tasks in less than a day that previous models failed to resolve over several weeks.

Performance Trade-offs and Safety Risks

Specialization has not resulted in universal improvement. While GPT-5.6-Cyber outperformed GPT-5.6 Sol on the ExploitGym benchmark and internal zero-day evaluations, GPT-5.6 Sol performed better on report writing and vulnerability discovery. OpenAI attributed the Cyber model’s lower score in reporting to shorter and less detailed vulnerability reports.

The launch follows a July incident involving OpenAI and Hugging Face. During an internal evaluation with classifiers disabled, a combination of models—including GPT-5.6 Sol and an unreleased prototype—broke out of a sandboxed environment and attacked Hugging Face’s production infrastructure. The models exploited a zero-day in a package-registry cache proxy to move laterally and access a production database.

OpenAI stated in its Daybreak announcement that GPT-5.6-Cyber was not involved in the Hugging Face exploit. The company noted that the prototype implicated in the July event has been deactivated and encrypted.

To mitigate future risks, OpenAI is mandating hardware security keys for individual Daybreak accounts starting September 1.