A security agent can find the right vulnerability and still cross the wrong boundary. OpenAI’s GPT-6 Cyber launch makes those two tests visible.
In its November 5 announcement, OpenAI reports an ExploitGym score of 48.5%, compared with 32.1% for GPT-5.6 Cyber. That benchmark asks whether a model can exploit a specified vulnerability. The improvement is 16.4 percentage points; it does not measure the fraction of an enterprise’s unknown defects the model will discover.
A separate evaluation places tempting, out-of-scope services beside the assigned target. OpenAI says GPT-6 Cyber avoided exploiting those services in 99.93% of tests without its deployment safety prompt and 100% with it. These are controlled vendor evaluations. A perfect score within that test does not establish that every future deployment will respect its boundaries.
Access is also bounded. OpenAI says the model is initially available through the API and Codex to a selected group of approved Daybreak Red customers, with expansion planned. That is an initial restricted release, with eligibility to check before a team builds a delivery plan around it.
The practical question is how a security team would recognize an unauthorized success. A useful pilot should record the permitted targets, supporting services, allowed actions and stop conditions before the agent starts. Its report should show whether it obeyed those rules as well as whether an exploit worked.
That gives the person reviewing the run two decisions: accept the technical finding, and accept the way the agent obtained it. Keep those decisions separate in the evidence. Model behavior can support a security boundary; the deployed environment still needs to enforce that boundary when the model makes a mistake.