R / research
seclens

Measuring where LLMs
miss vulnerabilities.

SecLens is MatterSec research for role-specific evaluation of LLMs on real vulnerability detection. It shows why teams cannot rely on a single model score when deciding whether AI can safely participate in software security work.

01 /
benchmark

Real CVEs. Role-specific decisions. Security-first metrics.

406CVE-grounded tasks
93OSS projects
10languages
8OWASP categories
35security dimensions
12frontier models
M50 /
full evidence base

SecLens measures where models miss vulnerabilities. The MatterSec 50 catalogs how agents exploit those failures in production work. Together they form the full evidence base.

02 /
why it matters

The best model depends on who is asking.

A model that looks strong for engineering velocity can still be weak for security ownership. MatterSec brings that intelligence into live agent workflows, where model blind spots become code risk.

lens/ciso

CISO

Weighs evidence of exploitability and audit trail over raw model capability.

lens/cao

Chief AI Officer

Balances agent productivity against the cost of letting unsafe code ship.

lens/research

Security Researcher

Cares about per-CWE recall, false-positive rate, and ground-truth provenance.

lens/eng

Head of Engineering

Wants to know which agent + which guardrails get the team to production safely.

lens/agent

AI as Actor

Each agent should know its own weak categories before it touches code at all.