OpenAI Launches GPT-5.6-Cyber Model Through Daybreak to Strengthen Cyber Defence

OpenAI has expanded its Daybreak cybersecurity programme by introducing two access tiers alongside the GPT‑5.6‑Cyber model. The GPT‑5.6‑Cyber model is available through Daybreak Red, and it is trained to offer improved performance in specialised cybersecurity tasks. The ChatGPT maker says the GPT‑5.6‑Cyber helps trusted defenders conduct legitimate security activities. The company said the GPT-5.6-Cyber has been used to discover previously unknown vulnerabilities in V8. OpenAI says GPT-5.6 Cyber is only available to select companies. 

OpenAI's GPT-5.6-Cyber Targets Vulnerability Discovery

OpenAI, in a blog post on Monday, announced the GPT‑5.6‑Cyber model, available through Daybreak Red access. It is designed to handle advanced cybersecurity tasks with fewer refusals. The company states that it has conducted an internal test covering tasks such as exploit-chain development, privilege escalation, and other advanced cybersecurity scenarios. OpenAI says it is better at tasks such as finding vulnerabilities, validating exploits, finding zero-day flaws and refusing fewer legitimate cybersecurity requests than its general-purpose models.

GPT-5.6-Cyber claimed to have completed 95 percent of these requests, compared with 1.5 percent for GPT-5.6 Sol and 2 percent for GPT-5.6 Sol through Daybreak Blue. OpenAI says it performed better than GPT-5.5-Cyber, which completed 57.3 percent of the requests.

GPT-5.6-Cyber is said to be trained for exploit development and advanced security research. On ExploitGym, GPT-5.6-Cyber reportedly outperformed both GPT-5.6 Sol and GPT-5.5-Cyber.

OpenAI says it tested the model's ability to identify novel zero-day vulnerabilities. GPT-5.6-Cyber claimed to have outperformed GPT‑5.6 Sol (Daybreak Blue) in identifying vulnerabilities of an open-source project, demonstrating their potential impact and offering a technical explanation.

Further, OpenAI says it tested GPT-5.6-Cyber on its internal Vulnerability Discovery and Report Writing evaluation to evaluate its ability to find security flaws in software and prepare detailed vulnerability reports. Both GPT-5.6 Sol and GPT-5.6-Cyber performed better than GPT-5.5-Cyber. GPT-5.6-Cyber scored lower than GPT-5.6 Sol on this evaluation. OpenAI said this may be because "the model sometimes produces shorter, less detailed vulnerability reports".

The company also evaluated the models on ExploitBench. The standard 300-turn limit, GPT-5.6 Sol through Daybreak Blue performed best and used fewer tokens. However, when the limit was increased to 600 turns, the performance gap between GPT-5.6 Sol and GPT-5.6-Cyber narrowed.

OpenAI claims that GPT-5.6-Cyber was used to investigate V8, the JavaScript engine used by Chrome, and it discovered two unknown vulnerabilities. OpenAI says its researchers validated the findings and reported them to Google through coordinated disclosure. Google fixed the vulnerability and assigned it CVE-2026-15903.

OpenAI said it gave GPT-5.6-Cyber early access to trusted customers, who used the model to improve their defensive cybersecurity workflows. The early-access partners included SpecterOps, SentinelOne and Palo Alto Networks.