OpenAI has launched GPT-5.6-Cyber, a cybersecurity-focused model designed for vulnerability research, penetration testing, and incident response. The model, built on GPT-5.6 Sol, is trained to improve performance on tasks like finding zero-day flaws and developing exploit chains, while reducing refusals for higher-risk, dual-use cyber activities.

The model is being offered through a new access tier called Daybreak Red, which gives vetted firms access to purpose-trained cybersecurity models for authorized vulnerability research, exploit validation, and security testing. A second tier, Daybreak Blue, provides access to frontier general-purpose models like GPT-5.6 Sol with built-in guardrails for defensive work.

OpenAI reports that GPT-5.6-Cyber completes 95.0% of advanced cybersecurity requests in an internal evaluation, compared to 1.5% for GPT-5.6 Sol and 2.0% with Daybreak Blue access. It also outperforms the earlier GPT-5.5-Cyber, which completed 57.3% of such requests. The model also showed gains on an ExploitGym benchmark, though it produced shorter, less detailed vulnerability reports than GPT-5.6 Sol in open-ended tasks.

The model identified several vulnerabilities, including CVE-2026-15903, a high-severity out-of-bounds read and write flaw in the V8 JavaScript engine that could allow remote code execution inside a sandbox. That flaw was patched by Google in mid-July 2026. Other discoveries include at least five flaws in a mobile operating system, three critical issues in a popular database, and over 400 privilege-escalation vulnerabilities in a common operating system kernel.

OpenAI has made the model available to partners like Accenture, Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, IBM, Palo Alto Networks, PwC, and Sophos. The company argues that reducing safeguards is necessary to close the "defense gap," but acknowledges the risks of misuse or misalignment.

Industry research cited in the announcement shows that AI models still struggle with patching: only 26.0% of generated patches fully resolved a vulnerability without changing behavior, while 53.9% failed to fix the issue, introduced new bugs, or both. The report also notes that attackers are using AI to speed up exploit development and broaden their targeting of newly disclosed flaws.

Source: The Hackers News