GPT-6 Cyber capabilities: The new era of automated security

On November 5, 2026, OpenAI launched GPT-6 Cyber to select customers in their Daybreak Red program. It arrived right alongside other new agent models like SWE-2 Flash. According to OpenAI and Reuters, this specialized variant of GPT-6 Astra is designed strictly for authorized security work like penetration testing and exploit development.

The new GPT-6 Cyber capabilities represent a massive jump in automated threat modeling. It hit 48.5% on the ExploitGym benchmark, which tests an agent's ability to turn known vulnerabilities into functional exploits. That is a 51% relative improvement over GPT-5.6 Cyber.

The model does not just pass tests; it writes functional exploits for real software. Reports from AI/TLDR and Reuters show GPT-6 Cyber autonomously discovered 948 local privilege escalation exploits in Linux. It also found 81 previously unknown high-severity bugs in Mozilla Firefox, plus a critical Brotli vulnerability (CVE-2026-111497) with a CVSS score of 9.2.

Finding a critical flaw in a widely used compression library proves these agents can handle complex codebases. This level of automated bug hunting forces teams to rethink how quickly they patch. We recently looked at how models now write code to attack Firefox, and this release accelerates that timeline.

Handing an AI the ability to write zero-day exploits requires strict boundaries. OpenAI classified the model as Critical under its Preparedness Framework, an upgrade from the previous High rating. During alignment evaluations, the model avoided exploiting out-of-scope services 99.93% of the time without deployment safety prompts, and hit 100% when those prompts were active.

Right now, access is limited to the API and Codex for authorized security work. If you manage to get access, you should estimate your API costs before running extensive penetration tests, as agent loops consume massive amounts of data.

The arrival of these models means attackers will soon have access to similar autonomous capabilities. Defensive teams need to automate their patching pipelines and assume that known vulnerabilities will be weaponized in hours, not weeks. We covered the basics of defending against offensive agents, but the speed of this new release requires immediate action.

While you review your security posture, you should also check your prompt sizes if you plan to feed large server log files into defensive AI tools. Start building internal guardrails now, because the era of autonomous security agents is officially here.