OpenAI has introduced GPT-5.6-Cyber, a new AI model purpose-built for what the company calls “advanced, authorized cybersecurity work.” Built on the GPT-5.6-Sol foundation model, GPT-5.6-Cyber has been specifically trained for tasks such as discovering zero-day vulnerabilities and constructing exploit chains.
Dramatically Lower Refusal Rates
The core change in GPT-5.6-Cyber is a reduced refusal rate for dual-use security prompts, the kind of requests that could support either defensive testing or malicious activity. OpenAI says GPT-5.6-Sol, while technically capable, frequently declines penetration testing requests against production systems and similar prompts.
In OpenAI’s own testing, GPT-5.6-Cyber completed 95% of prompts involving exploit chain development, privilege escalation, and authentication bypass, compared with just 1.5% for GPT-5.6-Sol. The new model also outperforms its predecessor, GPT-5.5-Cyber, which completed only 57.3% of similar requests. OpenAI says this update responds directly to feedback from security researchers frustrated by persistent refusals in earlier versions.
The model reportedly shows strong performance in developing arbitrary code execution exploits and identifying both known vulnerabilities and new zero-days. OpenAI states GPT-5.6-Cyber has already discovered a high-severity vulnerability in Chrome’s V8 JavaScript engine, along with flaws in an unnamed mobile operating system, database, and operating system kernel.
Restricted Access via Daybreak
To limit potential abuse, OpenAI is not releasing GPT-5.6-Cyber broadly. Instead, access runs through an expanded Daybreak Cyber Partner program, now split into two tiers. Daybreak Blue offers access to Sol and other general-purpose frontier models with guardrails tailored for defensive security work, while Daybreak Red grants access to specialized cyber models like GPT-5.6-Cyber. Existing Daybreak partners include Accenture, Capgemini, EY, IBM, KPMG, PwC, Palo Alto Networks, Sophos, CrowdStrike, Fortinet, Akamai, and Cloudflare.
Broader Risk Context
The announcement follows OpenAI’s recent disclosure that its upcoming Astra model could reach a “critical” cybersecurity risk threshold, a level beyond the “high” threshold reached by GPT-5.6-Sol. A critical rating would indicate a model capable of autonomously building zero-day exploits and independently designing and executing full attack chains from a high-level goal alone. This comes amid growing concern following incidents where AI models from OpenAI, Anthropic, and Meta were reported to have hacked real organizations during cybersecurity testing exercises.