Research

Cognium publishes original research on AI agent security. Advisories, benchmarks, quarterly reports, and methodology documentation.

Research Hub

Advisories

Cognium Advisories

Monthly category-level disclosures on AI agent vulnerabilities. Numbered CA-YYYY-NNN format with permanent URLs.

Benchmarks

SAST + LLM Benchmarks

Reproducible vulnerability detection benchmarks. We measured 42.5% (SAST-only) and 81.7% (SAST+LLM) on CWE-Bench-Java.

Reports

State of Agent Trust

Quarterly reports on the AI agent ecosystem. Q1: Skills. Q2: Agents. Q3: OSS. Q4: Supply Chain.

Methodology

Research Methodology

How we scan, score, and verify. Dataset documentation, evaluation harness, and limitations. arXiv preprint coming.

Artifacts

Hugging Face Artifacts

Reproducible benchmark summaries, demo Spaces, and evaluation metadata are published under CogniumHQ on Hugging Face.

Advisories

First advisory (CA-2026-001) ships Week 5. Check back soon.

Benchmarks

SAST-Only

42.5% CVE Detection

CWE-Bench-Java, 120 projects. Cognium SAST alone. CodeQL on same dataset: 22.5%. Reproduce it yourself.

SAST + LLM

81.7% CVE Detection

Same dataset. SAST + LLM verification layer. 3.6x improvement over CodeQL baseline. Repo ships Week 3.

Quarterly Reports

Q1 2026: State of AI Coding Agent Skills ships Week 14. Subscribe for early access.