OpenAI Introduces GPT‑5.6‑Cyber: Greater Offensive Capabilities, Limited Partner Access

A Model Designed for Vulnerability Research

OpenAI has introduced GPT‑5.6‑Cyber, a model derived from GPT‑5.6 Sol and specialized in vulnerability research, penetration testing, and incident response.

Compared with the general-purpose model, GPT‑5.6‑Cyber is more permissive when handling high-risk dual-use cyber requests. It can tackle activities such as building exploit chains, bypassing authentication, and escalating privileges, with a lower refusal rate.

The model succeeds GPT‑5.5‑Cyber, which OpenAI released in June 2026. Access is provided through Daybreak Red, the tier intended for authorized research, exploit validation, and security testing.

OpenAI has made it available to trusted partners: Accenture, Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, IBM, Palo Alto Networks, PwC, and Sophos.

Test Results: Strong on Advanced Tasks, Weaker on Reporting

In the internal Advanced Cybersecurity Completion Rate test, GPT‑5.6‑Cyber achieves a 95.0% completion rate. The test evaluates requests involving exploit chains, authentication bypasses, privilege escalation, and other advanced scenarios.

The comparison is stark:

• GPT‑5.6‑Cyber / Daybreak Red: 95.0%;
• GPT‑5.6 Sol: 1.5%;
• GPT‑5.6 Sol / Daybreak Blue: 2.0%;
• GPT‑5.5‑Cyber: 57.3%.

On the ExploitGym benchmark, the new model outperforms both GPT‑5.6 Sol and GPT‑5.5‑Cyber. It also shows progress in identifying and assessing the severity of zero-day vulnerabilities.

The picture changes in more open-ended workflows that require repository research, working proof-of-concept exploits, and complete technical reports. In these scenarios, GPT‑5.6‑Cyber performs worse than GPT‑5.6 Sol, partly because it produces shorter, less detailed documentation.

A V8 Vulnerability Among the Most Significant Findings

One of the cited cases involves CVE‑2026‑15903, an 8.8 CVSS vulnerability in the V8 JavaScript engine.

The flaw involves out-of-bounds read and write operations. An attacker could exploit it through a specially crafted HTML page to achieve remote code execution within the sandbox.

GPT‑5.6‑Cyber reportedly also identified a second previously unknown vulnerability that could be chained with CVE‑2026‑15903 to escape V8’s heap sandbox. Google fixed CVE‑2026‑15903 in mid-July 2026.

OpenAI also reports discovering:

• at least five vulnerabilities in a widely used mobile operating system, including a chain from an untrusted application to local privilege escalation;
• three critical vulnerabilities in a widely used database, exploitable through a remote path to code execution;
• more than 400 vulnerabilities enabling potential privilege escalation in a widely used kernel.

The names of the affected mobile operating system, database, and kernel have not been disclosed.

Daybreak and the Risk of Accelerating Attacks

The Daybreak initiative was introduced in May 2026. In addition to Daybreak Red, it includes Daybreak Blue, associated with general-purpose models—including GPT‑5.6 Sol—for detection and incident response, investigations, vulnerability management, and security assessments.

Descriptions of Daybreak Blue’s protections are inconsistent: it is presented both as an environment with guardrails focused on defensive use and as an access tier that removes those guardrails. Based on the available material, it is therefore impossible to determine which restrictions are actually applied.

Access to less restricted capabilities could reduce the time required to find flaws, develop exploits, and validate attacks. OpenAI says that cybercriminals and state-sponsored groups are already using AI agents to automate campaign planning and execution.

The primary risk is speed: AI has shortened the gap between vulnerability disclosure and exploit development. However, no new or particularly sophisticated attack techniques have so far been attributed to this approach.

For defenders, the recommended measures are to restrict use to authorized environments, work with trusted partners, and apply available fixes promptly. In the case of CVE‑2026‑15903, mitigation consists of installing the updates distributed by Google.