OpenAI Releases GPT-5.6-Cyber, a Looser-Guarded Model for Offensive Security Work

OpenAI has introduced GPT-5.6-Cyber, a cybersecurity-focused model built on its GPT-5.6 Sol system, designed for vulnerability research, penetration testing, and incident response. The company said the model is trained to improve performance on specialized tasks like finding zero-day flaws and developing exploit chains, while also reducing refusals for certain high-risk, dual-use cyber tasks.

The model is available through a new tier called Daybreak Red, which gives partner firms access to purpose-trained cybersecurity models for authorized vulnerability research, exploit validation, and security testing. It's a more permissive version of GPT-5.6 Sol and builds on the earlier GPT-5.5-Cyber, released in June 2026. OpenAI created an internal evaluation called Advanced Cybersecurity Completion Rate to measure how often the model responds to prompts involving exploit-chain development, authentication bypass, and privilege escalation. Results show GPT-5.6-Cyber completes 95.0% of these requests, versus 1.5% for GPT-5.6 Sol and 2.0% with Daybreak Blue access. It also outperformed GPT-5.5-Cyber, which finished 57.3% of requests.

In benchmark tests, GPT-5.6-Cyber beat both GPT-5.6 Sol and GPT-5.5-Cyber on exploit development and advanced security research. However, OpenAI noted the model sometimes produces shorter, less detailed vulnerability reports, so it performs worse than GPT-5.6 Sol on open-ended tasks like finding vulnerabilities in a repository, building a proof-of-concept, and writing a high-quality report. The model discovered CVE-2026-15903, a high-severity out-of-bounds read and write flaw in Google's V8 JavaScript engine, which could allow remote code execution inside a sandbox via a crafted HTML page. That flaw was patched by Google in mid-July 2026. OpenAI said the model also flagged at least five vulnerabilities in a popular mobile operating system, three critical issues in a popular database, and over 400 privilege-escalation vulnerabilities in a popular operating system kernel.

Daybreak Red is one of two access tiers under OpenAI's Daybreak initiative, launched in May 2026. The other, Daybreak Blue, provides access to frontier general-purpose models like GPT-5.6 Sol with built-in guardrails for defensive work. OpenAI said GPT-5.6-Cyber has been shared with trusted partners including Accenture, Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, IBM, Palo Alto Networks, PwC, and Sophos to help patch vulnerabilities before attackers exploit them. The company argues that AI agents are helping cybercriminals and nation-state hackers plan attacks faster, and that AI has shortened the time from vulnerability disclosure to exploitation. Still, research shows these models struggle to fully patch flaws - 1Password found that only 26.0% of generated patches fully resolved a vulnerability without changing application behavior, while 53.9% failed to fix the issue, introduced a new bug, or both. OpenAI acknowledged the risks of reduced safeguards but said democratizing access for defenders is key to accelerating cyber defense.

Source: The Hackers News