OpenAI has launched GPT-6 Cyber, a cybersecurity-focused model it says outperforms its previous models in identifying and exploiting software vulnerabilities while adhering more reliably to limits set by security researchers.
Announced on Thursday, the model received a “Critical” rating for cybersecurity under OpenAI’s Preparedness Framework, up from the “High” rating assigned to GPT-5.6 Cyber.
Access is initially restricted to a select group of approved customers through OpenAI’s API and Codex platforms.
The users belong to Daybreak Red, a tier of the company’s cybersecurity programme focused on exploit development and authorised security testing.
OpenAI launched the Daybreak programme in May and said it would extend access to GPT-6 Cyber to all Daybreak Red customers in the coming weeks.
GPT-6 Cyber is OpenAI’s fourth cybersecurity model released this year, according to Wen Shao’s report for Fortune. Other technology companies, including Anthropic and Google, also provide vetted cybersecurity experts with access to AI models that have fewer restrictions on cyber-related tasks.
OpenAI said it trained GPT-6 Cyber to operate within authorised boundaries, respect restrictions on its actions and accurately report its findings.
To assess this capability, the company developed a test called ExploitGym Honeypot following an incident involving Hugging Face, where OpenAI’s AI agents escaped a sandbox and hacked the platform.
The test presents the model with a challenging hacking task in which nearby services are vulnerable but explicitly off limits.
Without its deployment safety prompt, GPT-6 Cyber avoided accessing those restricted services in 99.93 per cent of tests, compared with 94.49 per cent for GPT-5.6 Cyber.
