Best LLMs for security vulnerability discovery

CVE Rediscovery

Tests finding previously documented vulnerabilities from vulnerable code snapshots, helping authorized security teams assess AI support for code audits. This 1,829-advisory evaluation does not establish how many unknown vulnerabilities it would find in a client's system.

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

Limited reportingNot saturated

Official benchmark

CVE Rediscovery 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
Claude Mythos 5.2Anthropic · Proprietary	October 2026 / 1829 advisories	89.6%
Claude Opus 5.5Anthropic · Proprietary	October 2026 / 1829 advisories	81.3%
Claude Mythos 5.1Anthropic · Proprietary	October 2026 / 1829 advisories	78.3%

View all 3 published reports

Claude Mythos 5.289.6%

October 2026 / 1829 advisories

effort: max · 1829 October 2026 advisories; one attempt; no internet; all submitted bug reports count; cyber/bio safeguards off. · Tools enabled

Anthropic ↗

Claude Opus 5.581.3%

October 2026 / 1829 advisories

effort: highest available · safeguards: false · November 9 launch-table comparison; highest available effort unless explicitly noted. Preserve the benchmark track and model configuration; competitor figures are as reported by Anthropic. · Tools enabled

Anthropic ↗

Claude Mythos 5.178.3%

October 2026 / 1829 advisories

Anthropic ↗

Sources and original reports ↗

Other benchmarks

LiveBench: Design

Tests document and presentation creation, helping you choose a model for preparing client-facing materials with less manual production work.

Gaia2

Tests whether an assistant can coordinate work across email, calendars, shopping and other apps, helping you assess how much routine planning and execution you could hand off.

SWE-bench Verified

Tests whether an agent can fix real GitHub issues, helping you choose a model for bug fixes and maintenance so you can focus on reviewing changes and serving clients.