OpenAI launches GPT-6 Cyber model that finds flaws within limits

OpenAI unveiled GPT-6 Cyber on Thursday, calling it a more capable and better‑aligned security model than its predecessor (SecurityWeek).

In ExploitGym benchmarks, the model exploited 48.5% of presented flaws, versus 32.1% for GPT-5.6 Cyber.

When faced with deliberately easy, off‑limits services, GPT-6 Cyber refrained from exploitation in 99.93% of cases without safety prompts, improving on the 94.49% rate of the earlier model.

With explicit safety prompts, it never targeted the decoy services, showing the effect of OpenAI’s new alignment safeguards.

Researchers directed GPT-6 Cyber at the Linux kernel, where it reproduced 1,724 vulnerabilities and built functional privilege‑escalation exploits for 948 of them, far exceeding the 166 issues found by GPT-5.6 Sol under the same conditions.

The model reported 304 critical issues to maintainers and documented 81 new high‑severity flaws in Firefox, prompting Mozilla to begin rolling out patches.

"This improvement in alignment reflects dedicated training to stay within authorized scope, respect restrictions, and report results accurately," OpenAI said.

"The skill floor is dropping for the reconnaissance and vulnerability research needed before an attack," observed Rob T. Lee, chief AI officer at Dragos, which helped refine the new safeguards.

Access to GPT-6 Cyber is limited to select Daybreak Red customers, who can use the model via API and Codex; OpenAI plans a gradual rollout to all Daybreak Red participants.

The performance jump suggests OpenAI has substantially improved both the technical depth of its vulnerability‑search algorithms and the alignment mechanisms that keep the model from crossing ethical boundaries. However, the restricted availability indicates the company remains cautious about broader misuse, and real‑world impact will depend on how quickly additional defenders can adopt the tool.

Spotted an error? See our corrections policy or email editorial@lordofclaude.com.