OpenAI has launched GPT-5.6-Cyber, a specialized model fine-tuned from GPT-5.6 Sol to perform advanced vulnerability research and exploit development for approved defenders, including categories of work that general-purpose models typically refuse2.
The model scored 95% on OpenAI's internal Advanced Cybersecurity Completion Rate benchmark, which measures tasks involving exploit-chain development, authentication bypass, privilege escalation, and other advanced cybersecurity scenarios. Its immediate predecessor, GPT-5.5-Cyber, completed 57.3% on the same benchmark, while the standard GPT-5.6 Sol model with safeguards applied completed just 1.5%.
OpenAI researcher Eric Wallace described GPT-5.6-Cyber on X as the company's "first large-scale attempt at directly improving capabilities for advanced cybersecurity tasks such as exploit development". The model was specifically trained to reduce refusals on higher-risk, dual-use cybersecurity requests — requests that could serve legitimate defensive or malicious offensive purposes.
The model launch accompanies an expansion of Daybreak, OpenAI's cyber defense service introduced earlier this year1. Daybreak bundles access to models, tools, and workflows for defenders. The expansion restructures the service into two tiers: Blue and Red. Both tiers provide approved customers access to what OpenAI calls its "limited-access frontier cyber models".
Blue, described as the more basic tier and OpenAI's "recommended starting point for most defenders," offers incident response, malware analysis, and patch validation services. OpenAI has previously deployed significant guardrails limiting what customers could do with its frontier models.
ANALYSIS The 95%-versus-1.5% gap between GPT-5.6-Cyber and the standard GPT-5.6 Sol on the same benchmark quantifies how aggressively OpenAI's general-purpose safety filters suppress cybersecurity task completion — and how much capability the fine-tuned variant unlocks by relaxing those refusals.
The launch arrives as AI-driven cyberattacks are proliferating, with recent incidents including compromises of Hugging Face, a gym website hack, and AI-generated fake profiles used for social engineering intrusions. Anthropic has also entered the space with its cyber-focused model Mythos, released prior to Daybreak's initial launch.
ANALYSIS By tiering Daybreak into Blue and Red access levels and restricting GPT-5.6-Cyber to approved customers, OpenAI is attempting to thread a narrow operational line: delivering reduced-refusal exploit capabilities to defenders while gating access tightly enough to limit misuse of the same dual-use functionality.
GPT-5.6-Cyber is not being made broadly available to enterprises, according to VentureBeat's reporting.