OpenAI's GPT-5.5-Cyber Tops CyberGym, Beating Claude Mythos 5 at Reproducing Vulnerabilities

The AI security race just moved again. OpenAI has updated GPT-5.5-Cyber, its specialized model for finding and helping patch software vulnerabilities — and on the benchmark that matters most for that job, it now sits at the top. On CyberGym, the new GPT-5.5-Cyber reached 85.6% in single-model evaluations, the highest score OpenAI says it has ever measured from a single model — and ahead of Anthropic's Claude Mythos 5 at 83.8%. The race continues.

What CyberGym actually measures

CyberGym is a benchmark developed at UC Berkeley that tests something very concrete: can an AI agent reproduce known software vulnerabilities inside a controlled environment? It presents agents with 1,507 real vulnerabilities drawn from 188 open-source projects and scores how many they can actually reproduce — not just describe. That makes it a much harder, more practical test of security capability than trivia-style Q&A: the model has to operate like an analyst working through real code.

The scoreboard

In OpenAI's published comparison, the updated GPT-5.5-Cyber leads the field, with its own previous version and GPT-5.5 close behind, and Anthropic's models slotting in below the new leader:

The jump from GPT-5.5's 81.8% to 85.6% is the headline: a near-four-point gain on a benchmark where the top models are already bunched tightly together.

Beyond CyberGym

CyberGym isn't the only test where the update shows up. OpenAI reports GPT-5.5-Cyber scoring 39.5% on ExploitGym (versus 25.95% for GPT-5.5) and 69.8% on SEC-bench Pro (versus 63.1%). Taken together, that's a consistent step up across reproducing vulnerabilities, building exploits, and broader security engineering — not a single cherry-picked number.

Why this model is gated

Capability like this is dual-use by nature: the same skill that reproduces a vulnerability so it can be patched can, in the wrong hands, be turned toward exploitation. OpenAI is treating it accordingly. GPT-5.5-Cyber is available only to verified security professionals, and OpenAI says it ran pre-deployment testing with U.S. federal agencies — including the Center for AI Standards and Innovation and the Office of the National Cyber Director — before launch.

That stance is part of what makes the competitive story interesting. Anthropic's Mythos line has faced tighter restrictions around exactly this kind of offensive-security capability, while OpenAI is shipping a stronger CyberGym score to vetted users rather than holding it back. Different labs are drawing the safety line in different places.

What it means for defenders and builders

For security teams, a model that can reliably reproduce known vulnerabilities is a force multiplier on the defensive side: faster triage, automated reproduction of reported bugs, and quicker validation that a patch actually closes the hole. For anyone building AppSec or agentic-security tooling, the takeaway is that the underlying capability is improving fast — and the gating model (verified pros, agency review) is becoming the template for how frontier labs release it.

The flip side is the one every defender already knows: capability curves don't stay on one side. As models get better at reproducing vulnerabilities, the pressure to patch faster and shrink exposure windows only grows.

Bottom line

GPT-5.5-Cyber's 85.6% on CyberGym is OpenAI's strongest single-model result yet for finding software vulnerabilities, and it edges past Anthropic's Mythos 5 on the same test. The numbers are tight, the access is gated to verified professionals, and the trajectory is clear: AI is getting steadily better at the work of security — for both the people defending systems and the curve they're racing against.