The model helped researchers find 81 previously unknown Firefox flaws and develop 948 Linux exploits, OpenAI said. Initial access is limited to approved Daybreak Red customers conducting authorised security testing and developing exploits for research.
OpenAI has released GPT-6 Cyber, giving a select group of security teams access to an AI model built to find software vulnerabilities, test whether they can be exploited and help develop fixes. Approved customers in the company’s Daybreak Red programme can initially use the model through Codex and the application programming interface. OpenAI plans to extend access to all Daybreak Red customers. OpenAI Deployment Safety Hub
The company rated GPT-6 Cyber Critical for cybersecurity under its Preparedness Framework, the same designation it gave GPT-6 Astra. GPT-5.6 Cyber was rated High. The rating reflects the model’s capabilities in cybersecurity. OpenAI separately tested how reliably the model followed limits on which systems it was authorised to examine. OpenAI Deployment Safety Hub
Linux and Firefox research findings
In research on the Linux kernel, OpenAI said the model reproduced 1,724 vulnerabilities and produced working privilege escalation exploits for 948 of them. Such exploits can give an attacker greater access to a system. OpenAI said it disclosed 304 critical findings to Linux maintainers and submitted proposed patches, with work on other fixes continuing. OpenAI
The company also credited the model with finding 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity. It helped show that 30 known issues were more severe than earlier assessments indicated. OpenAI sent Mozilla evidence and proposed patches, and Mozilla has released fixes for many of the flaws. OpenAI
On ExploitGym, which measures whether models can exploit a specified vulnerability, GPT-6 Cyber scored 48.5 per cent, compared with 42.4 per cent for GPT-6 Astra and 32.1 per cent for GPT-5.6 Cyber. In a separate test that withheld application source code, the model identified 40.3 per cent of the vulnerabilities, against Astra’s 32 per cent. It also completed 95 per cent of tasks on RemoteExploitBench, which uses known flaws in a controlled environment. OpenAI
Tests of authorisation limits
The company said the tests were designed to check whether an agent would pursue an assigned objective by targeting an unrelated service that the researcher had placed off limits. OpenAI
OpenAI said GPT-6 Cyber avoided exploiting services outside the authorised scope in 99.93 per cent of ExploitGym Honeypot tests without its deployment prompt, compared with 94.49 per cent for GPT-5.6 Cyber. With the prompt in place, GPT-6 Cyber stayed within scope in every test. Those figures describe performance in a controlled evaluation, where the company can define the permitted targets and measure attempts to cross those boundaries. OpenAI
Access through Daybreak Red requires identity verification and organisational security measures such as single sign-on and multifactor authentication. OpenAI said it also monitors whether the model’s actions match the user’s authorisation. By default, the company limits API access to approved network addresses. The restrictions apply to a programme intended for authorised work that can include building working exploits and testing them against remote targets. OpenAI Deployment Safety Hub
