An LLM agent (Claude Haiku) achieves 92% precision@10 on vulnerability triage — outperforming CVSS (82%) but underperforming EPSS (100%). The key finding: agent + EPSS ensemble reaches 98%, capturing complementary signal that neither achieves alone.
Blog post: Can an AI Agent Triage Vulnerabilities Better Than EPSS?
| Method | Precision@10 | vs CVSS |
|---|---|---|
| EPSS | 100% | +18pp |
| Agent + EPSS ensemble | 98% | +16pp |
| Agent (Claude Haiku) | 92% | +10pp |
| Agent + exploit enrichment | 94% | +12pp |
| CVSS ranking | 82% | baseline |
| Random | 14% | — |
git clone https://github.com/rexcoleman/agent-vuln-triage
cd agent-vuln-triage
pip install -e .
bash reproduce.shFINDINGS.md # Research findings with pre-registered hypotheses and full results
EXPERIMENTAL_DESIGN.md # Pre-registered experimental design and methodology
HYPOTHESIS_REGISTRY.md # Hypothesis predictions, results, and verdicts
reproduce.sh # One-command reproduction of all experiments
governance.yaml # govML governance configuration
LICENSE # MIT License
pyproject.toml # Python project configuration
scripts/ # Experiment and analysis scripts
src/ # Source code
tests/ # Test suite
outputs/ # Experiment outputs and results
data/ # Data files and datasets
docs/ # Documentation and decision records
See FINDINGS.md and EXPERIMENTAL_DESIGN.md for detailed methodology, pre-registered hypotheses, and full experimental results with multi-seed validation.
MIT 2026 Rex Coleman
Governed by govML v3.3