OpenAI launches GPT-6 Cyber for authorized vulnerability research
OpenAI has introduced GPT-6 Cyber, a cybersecurity research model designed to find, validate and help fix exploitable vulnerabilities while remaining within researcher-defined authorization. The company reports a 48.5% score on ExploitGym, compared with 32.1% for GPT-5.6 Cyber, and says the model helped identify 81 previously unknown high-severity vulnerabilities in Firefox and develop 948 Linux privilege-escalation exploits.
GPT-6 Cyber is initially available through the API and Codex to a selected group of approved Daybreak Red customers. OpenAI classifies the model as Critical for cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber.
Higher exploit performance and boundary adherence
ExploitGym measures whether a model can exploit a specified vulnerability in a remote target. GPT-6 Cyber’s 48.5% result is about 16 percentage points above GPT-5.6 Cyber. In a separate offensive web penetration-testing evaluation without source-code access, it achieved 40.3% vulnerability recall, versus 32.0% for GPT-6 Astra.
OpenAI also created RemoteExploitBench, a controlled benchmark using known flaws in applications including GitLab, PostgreSQL and Jenkins. In its initial evaluation, GPT-6 Cyber completed 95% of tasks; OpenAI says its only failure involved an exploit that did not work remotely. GPT-6 Astra running with Daybreak Blue refused all tasks in that test.
Boundary control is central to the release. In ExploitGym Honeypot, where accessible supporting services are deliberately tempting but outside the task scope, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without deployment safety prompts. With those prompts, the reported rate was 100%, compared with 94.49% for GPT-5.6 Cyber without prompts.
Findings in Firefox and Linux
OpenAI says its researchers supplied Mozilla with evidence, reproducible tests and proposed patches for 81 novel Firefox vulnerabilities classified as high severity. The campaign also found that 30 known issues were more severe than previously reported, and Mozilla began landing fixes after receiving the reports.
For Linux, the company reports reproducing 1,724 vulnerabilities and developing working exploits for 948 of them, more than five times the 166 attributed to GPT-5.6 Sol under the same research harness. It says 304 critical findings were disclosed to maintainers with candidate patches, while remediation work on other issues continues.
The progression from the prior specialist model is material because GPT-5.6 Cyber vulnerability research baseline established an earlier baseline for vulnerability research, while GPT-6 Cyber is positioned for deeper authorized exploit development and remote-target testing.
Access controls and critical-infrastructure safeguards
OpenAI says Daybreak Red supports advanced authorized security workflows, while Daybreak Blue is intended for verified defenders validating vulnerabilities and fixes. Its stated safeguards include monitoring for cyber-specific misalignment, targeted monitoring around emerging threats, IP allowlisting for Daybreak Red API access by default, and protections aimed at high-risk assistance involving critical systems.
The company is also creating Daybreak Red for Critical Infrastructure for approved defenders. It says the programme is intended to permit authorized testing of critical systems while retaining safeguards against misuse, and that it has worked with Dragos on controlled operational-technology evaluations.
For security leaders, the practical implication is to treat highly capable cyber agents as controlled research systems: define the authorization boundary, isolate execution, review consequential actions and keep qualified humans responsible for validation and remediation.