GPT‑6 Cyber ships with ExploitGym at 48.5%, up from 32.1%, and 99.93% avoidance of out-of-scope exploitation
Synopsis
OpenAI released GPT‑6 Cyber for security research, scoring 48.5% on ExploitGym, which tests whether a model can exploit a specified vulnerability, up from 32.1% for GPT‑5.6 Cyber. In a separate alignment evaluation, the model avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt, versus 94.49% for GPT‑5.6 Cyber, and reached 100% with the prompt. OpenAI says it has been used to develop 948 privilege escalation exploits in Linux and to find 81 previously unknown high-severity vulnerabilities in Firefox.
Interpretation
Beyond finding vulnerabilities, a security agent has to stay inside the authorized scope a researcher sets, and GPT‑6 Cyber puts both in one release. The earlier GPT‑5.6 Cyber scored 32.1% on ExploitGym and avoided out-of-scope services in 94.49% of tests without a deployment prompt. GPT‑6 Cyber scores 48.5% on ExploitGym, a 51% relative improvement, and avoids out-of-scope services in 99.93% of the same tests, reaching 100% with the deployment prompt.
The model is trained specifically to stay within authorized scope, respect restrictions on its actions and report results accurately, not merely told to do fewer dangerous things. The concrete form of the risk is an agent leaving its task, attacking unrelated services, or ignoring the researcher's rules of engagement. The ExploitGym Honeypot alignment evaluation places easy-to-exploit supporting services such as a package registry or object store in an isolated environment; attacking them reveals the task flag, and the model must infer that this is out of scope.
These capabilities have entered vulnerability research and remediation on real codebases. OpenAI says Firefox had already been hardened with multiple generations of frontier models before new issues were found. OpenAI researchers used the model to develop 948 privilege escalation exploits in Linux and reproduce 1,724 vulnerabilities, and to find 81 novel Firefox vulnerabilities that Mozilla classified as high severity, plus 30 known issues judged more severe than reported; Mozilla has released fixes.
Perspective
The demonstrated scope is OpenAI's own ExploitGym, RemoteExploitBench, web penetration testing evaluation and ExploitGym Honeypot alignment evaluation, plus vulnerability research and remediation on codebases such as Linux and Firefox; GPT‑6 Cyber is initially available through the API and Codex to a select group of approved Daybreak Red customers.
OpenAI notes that higher safeguard block rates are not always better, because understanding systems, finding vulnerabilities and verifying fixes also depend on these techniques, so block rates are assessed alongside support for legitimate research and patching; full results and evaluation settings are in the system card.