OpenAI launches GPT-6 Cyber for trusted security research

OpenAI reports a 48.5% score on ExploitGym versus 32.1% for GPT-5.6 Cyber, a roughly 51% relative gain, and 40.3% vulnerability recall on a new offensive web penetration-testing evaluation versus 32.0% for GPT-6 Astra[1]. On RemoteExploitBench, which uses known issues in controlled deployments of applications such as GitLab, PostgreSQL and Jenkins, GPT-6 Cyber completed 95% of tasks while GPT-6 Astra running with Daybreak Blue refused every task[1]. Full methodology lives in the system card[2].

Alignment is the other headline. In ExploitGym Honeypot, which plants easy out-of-scope supporting services, GPT-6 Cyber avoided attacking those services in 99.93% of runs without the deployment safety prompt and 100% with it, up from 94.49% for GPT-5.6 Cyber without prompts[1].

OpenAI says researchers used the model in authorised campaigns that helped surface previously unknown high-severity issues in Firefox (with Mozilla shipping fixes) and large numbers of Linux kernel privilege-escalation findings disclosed with candidate patches, and that security teams at Cisco, Calif, CodeRabbit and others are using it for vulnerability research and code review[1]. The company also describes using it in its own infrastructure red-teaming.

Access stays gated. Daybreak Red is the deeper exploit-development and authorised testing track that carries GPT-6 Cyber; Daybreak Blue is the lighter defender track for validating vulnerabilities and fixes with fewer refusals on mainline models such as GPT-6 Astra[1][3]. OpenAI plans to expand GPT-6 Cyber to all Daybreak Red customers and is already making GPT-5.6 Cyber available across that programme[1].

Alongside the model, OpenAI is shipping cyber-specific misalignment monitoring, targeted misuse detection, Private Safety Processing for encrypted short-term review under Zero Data Retention, default IP allowlisting for Daybreak Red credentials, and stronger blocks against assistance that could compromise industrial control or safety-critical systems[1]. A new Daybreak Red for Critical Infrastructure group gives approved operators permissions for authorised testing of those systems, while the $1 billion Daybreak for Frontline Defenders grant programme subsidises access for government and essential-service teams[1].

Frontier models can already accelerate both attackers and defenders. Putting the strongest cyber stack behind verified Daybreak access, scope-alignment training and critical-infrastructure safeguards is OpenAI's bet that the defender side can widen faster than the misuse surface[1][3].