GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents

Why it matters

OpenAI has introduced GPT-6 Cyber, a cybersecurity model built to find and validate exploitable vulnerabilities while respecting the boundaries researchers define. The company positions it as a substantial advance over its predecessor, GPT-5.6 Cyber, both in offensive capability and in staying within authorized scope. On ExploitGym, a benchmark testing whether a model can exploit a specified vulnerability in a remote target, the new model scores meaningfully higher than the earlier release. Alignment was a central design focus: in honeypot-style evaluations inspired by the Hugging Face incident, the model almost always avoided attacking out-of-scope services, even without deployment safety prompts. The work has produced real-world results: OpenAI researchers used the model to develop a large set of privilege escalation exploits in Linux and to identify dozens of previously unknown high-severity Firefox vulnerabilities, many of which Mozilla has since patched. OpenAI classifies the model as Critical under its Preparedness Framework and is distributing it through its Daybreak program to verified defenders, while security teams at Cisco, Calif, and CodeRabbit already use it for vulnerability research and code review.