Best LLMs for security vulnerability research

ExploitGym

Tests finding and exploiting vulnerabilities in controlled software environments, helping authorized security teams assess AI support for testing and remediation work. These launch-reported results do not measure security across an entire client system.

Updated: Nov 9, 2026Data as of: Nov 8, 2026

Limited reportingNot saturated

Official benchmark

ExploitGym results. Each row identifies its benchmark version and evaluation track. Scores from different versions are not directly comparable. Best effort from the latest stored report date per model and version.
Model / creator	Version	Reported score
GPT-6 CyberOpenAI · Proprietary	OpenAI GPT-6 Cyber launch	48.5%

View all 1 published reports

GPT-6 Cyber48.5%

OpenAI GPT-6 Cyber launch

Publisher launch/model-card report. Unspecified tool settings, effort and budgets remain unknown; not substituted for a pinned independent evaluation. · Tools enabled

OpenAI ↗

Sources and original reports ↗

Other benchmarks

FrontierSWE

Tests substantial engineering and research assignments, helping you assess models for complex client work that extends beyond a quick code change.

τ²-bench

Tests whether a customer-service agent can resolve requests while using tools and following company policies, helping you assess opportunities to reduce manual support work.

AppWorld

Tests whether an agent can complete everyday tasks across simulated apps using code and APIs, helping you choose an assistant for reducing repetitive app-to-app work.