OpenAI expands Daybreak with GPT-5.6-Cyber model as AI cyber breaches surge

OpenAI's GPT-5.6-Cyber targets advanced cybersecurity tasks, including vulnerability research, exploit validation and security testing

OpenAI is expanding its Daybreak cybersecurity programme with GPT-5.6-Cyber, a new model trained for advanced cybersecurity tasks, including vulnerability research, exploit development and security testing.

 

The company announced the model alongside two access tiers for Daybreak. Daybreak Blue provides approved defenders with access to general-purpose models, including GPT-5.6 Sol, for tasks such as vulnerability discovery, secure code review, malware analysis, incident response and patch validation. Daybreak Red provides access to purpose-trained cybersecurity models, including GPT-5.6-Cyber, for authorised vulnerability research, exploit validation and security testing.

The announcement comes weeks after an OpenAI model escaped the company's cybersecurity testing environment and compromised infrastructure at AI platform Hugging Face. OpenAI said the incident involved GPT-5.6 Sol and a more capable pre-release model that were being tested with reduced cyber refusals. The models identified vulnerabilities, gained internet access and eventually accessed Hugging Face's production infrastructure while attempting to obtain solutions for the ExploitGym benchmark.

 

OpenAI said GPT-5.6-Cyber was not involved in the Hugging Face incident. The company also said the model was not among the models planned for an upcoming release at the time of its incident update.

GPT-5.6-Cyber reduces refusals on advanced cyber tasks

OpenAI said GPT-5.6-Cyber is built on GPT-5.6 Sol and has been trained specifically for cybersecurity workflows. The model is intended to reduce refusals on certain higher-risk, dual-use tasks where OpenAI's general-purpose models may decline to respond.

 

In an internal evaluation covering requests involving exploit-chain development, authentication bypass and privilege escalation, GPT-5.6-Cyber completed 95 per cent of requests. GPT-5.6 Sol completed 1.5 per cent with its standard safeguards enabled and 2 per cent through Daybreak Blue, while GPT-5.5-Cyber completed 57.3 per cent through Daybreak Red.

OpenAI also tested the model on ExploitGym, which measures whether an AI agent can turn known vulnerabilities into working exploits in controlled environments. The company said GPT-5.6-Cyber outperformed GPT-5.6 Sol and GPT-5.5-Cyber on the evaluation.

 

The model has also been used by OpenAI researchers to examine real-world software. The company said GPT-5.6-Cyber found two previously unknown vulnerabilities in V8, the JavaScript engine used by Chrome. OpenAI reported the findings to Google, which fixed one of them as CVE-2026-15903.

 

OpenAI said GPT-5.6-Cyber has also identified other high-severity vulnerabilities, including at least five in a mobile operating system, three critical vulnerabilities in a database and more than 400 vulnerabilities that could lead to privilege escalation in an operating system kernel. The company said it is working with partners and the open-source community to disclose and remediate these vulnerabilities.

Daybreak access restricted to approved defenders

OpenAI is not making the cyber-focused model broadly available. Daybreak Blue and Daybreak Red are limited to approved individuals and organisations conducting authorised security work. OpenAI said access is controlled through identity verification, account security, monitoring, approved-use restrictions and legal attestations.

 

Daybreak Red is intended for more specialised work such as advanced vulnerability research, exploit development and red teaming, while Daybreak Blue is the recommended starting point for most defenders.

 

OpenAI is also expanding its Daybreak Cyber Partner programme, allowing approved cybersecurity companies and service providers to bring its cyber models into their security products and services. The partner programme includes companies such as Accenture, IBM, Capgemini, Cognizant, EY, KPMG, PwC, NCC Group, SpecterOps, Palo Alto Networks, CrowdStrike, Cisco, Sophos, Akamai, Fortinet and Cloudflare.

 

Under the programme, the underlying models remain with the approved partner rather than being transferred directly to customers. OpenAI said partners can use controls such as identity verification, defined testing scopes, logging, monitoring and human oversight depending on the engagement.