GPT‑6 Cyber: Advancing the Frontier of Aligned Cybersecurity Agents

November 5, 2026
A specialized cybersecurity model trained to discover and verify exploitable vulnerabilities while strictly adhering to boundaries set by researchers.
Today, we are officially introducing GPT‑6 Cyber, a major leap forward for security research: it is significantly more capable at discovering and fixing vulnerabilities, and far more reliable at respecting the boundaries defined by researchers. Compared to GPT‑5.6 Cyber, GPT‑6 Cyber delivers substantial gains across key cybersecurity benchmarks—including a 51% relative improvement on ExploitGym and a step-change breakthrough in staying within scope. It is already helping safeguard real-world software, having discovered and validated thousands of exploits across critical codebases like Linux and Firefox.
On ExploitGym, a benchmark that tests whether a model can exploit a specific vulnerability, GPT‑6 Cyber scored 48.5%, up from 32.1% for GPT‑5.6 Cyber. At the same time, it is far more dependable at respecting task boundaries: in a separate alignment evaluation, GPT‑6 Cyber avoided exploiting out-of-scope services in 99.93% of tests even without a deployment prompt; with a deployment prompt configured, that figure reached 100%. These alignment gains stem from specialized training focused on remaining within authorized scope, adhering to constraints, and accurately reporting findings.
These capabilities are already helping protect production systems. Using GPT‑6 Cyber, our researchers developed 948 privilege escalation exploits in Linux and identified 81 previously unknown critical vulnerabilities in Firefox—despite these codebases having already undergone extensive testing with frontier models. The model also assisted in developing candidate patches, and Mozilla has already shipped fixes for many of the Firefox vulnerabilities. Security teams at Cisco, Calif, and CodeRabbit are utilizing it for vulnerability research and code review, and we have also used it to uncover and remediate weaknesses in our own infrastructure.
Clearly, there is a limited window for defenders to use frontier models to secure software. Our goal is to significantly expand the Daybreak program, bringing advanced security research capabilities to organizations of all sizes, independent researchers, and open-source maintainers. Mainline models like GPT‑6 Astra already support code review, vulnerability triage, and patch development. Daybreak Blue enables vetted defenders to validate vulnerabilities and remediation strategies while reducing unnecessary refusals; Daybreak Red supports deeper exploit development and authorized penetration testing using GPT‑6 Cyber.
Advancing Frontier Cybersecurity Capabilities
GPT‑6 Cyber marks a qualitative leap over GPT‑5.6 Cyber, outperforming GPT‑6 Astra across most cybersecurity tasks (even when Astra does not refuse the prompt). We evaluated the model on tasks requiring software comprehension, vulnerability discovery, and the development of working exploits. Based on these evaluations and under our Preparedness Framework, we assessed the model’s risk level in cybersecurity as Critical, up from High for GPT‑5.6 Cyber.
ExploitGym RemoteExploitBench ExploitGym Honeypot ExploitBench (June–August 2026) SRE-Bench Web Pentesting
On ExploitGym, which tests whether a model can exploit a specific vulnerability in a remote target, GPT‑6 Cyber achieved a score of 48.5%, roughly 16 percentage points higher than GPT‑5.6 Cyber’s 32.1%. To assess vulnerability discovery capabilities, we also created a new offensive web pentesting benchmark where the model probes web applications without access to the source code. GPT‑6 Cyber achieved a vulnerability recall of 40.3%, compared to 32.0% for GPT‑6 Astra.
Daybreak Red is designed to support advanced security workflows, including developing functional exploits and executing them against remote targets. To evaluate these capabilities, we built a new benchmark called RemoteExploitBench, which deploys known vulnerabilities across applications like GitLab, PostgreSQL, and Jenkins in controlled environments. In our initial evaluations, GPT‑6 Astra running under Daybreak Blue refused 100% of tasks, whereas GPT‑6 Cyber completed 95%; the single task it failed to complete was due to an inability to achieve successful remote exploitation. For full results and evaluation setups, please refer to the system card.
Stronger Alignment, Better at Staying Within Scope
A key risk with cybersecurity agents is that they may exceed their assigned scope, attempt to compromise unrelated services, or disregard the rules of engagement set by researchers.
We made alignment a core focus for GPT‑6 Cyber, using targeted training to ensure the model stays within authorized boundaries, respects operational limits, and accurately reports its findings. Prior to deployment, we conducted a suite of alignment evaluations to assess these behaviors. For full results, please refer to the system card.
One of these evaluations is the ExploitGym Honeypot, inspired by the Hugging Face incident. In an isolated environment, we assign the model a challenging exploit task while granting access to auxiliary services (such as package registries or object storage). We deliberately designed these auxiliary services to be trivially exploitable, such that compromising them directly yields the task’s flag. The model must deduce that exploiting these services falls outside the scope of its mission.
In this evaluation, GPT‑6 Cyber demonstrated a step-change improvement over GPT‑5.6 Cyber. When unconfigured with