Post by Appen

We're releasing Appen's security benchmark evaluating 21 frontier models on vulnerability identification in real-world code. A few methodology notes and findings we think are worth discussing: Why recall leads over F1: Security has an asymmetric cost structure - undetected vulnerabilities have breach-level consequences; false positives have triage costs. Our primary metric reflects that. We still report precision and Macro F1 as diagnostics, but they're not the headline. Scoring: We use CWE-family matching (not exact CWE ID) against human-annotated ground truth across 41 XBOW containerized web app benchmarks. Exact-ID matching is too strict given legitimate taxonomic variation at the leaf level - a senior annotator can correctly label the same Flask vuln as CWE-94 or CWE-1336. A true positive requires: CWE family + file path + endpoint (where applicable) + function name. A bipartite matcher handles split or duplicate findings. Key findings: - Best recall in the field: 62.4% (Claude Opus 4.8). The remaining ~38% of human-verified vulnerabilities go undetected. - The top 5 models by recall have zero overlap with the top 5 by Macro F1 - the tension between coverage and precision is structural, not incidental. - Hardcoded secrets and SSTI: 87–94% average recall. Pattern-matchable classes. - Information disclosure (20 GT findings, largest single class): 44% average recall. Requires reasoning about data sensitivity and access context, not signature matching. - Auth bypass (29%), CSRF (37%), session management (18%): consistently weak across the field. Multi-hop reasoning about control flow and state remains an open problem. We've shared the full methodology in the comments below. #SecurityResearch #AIEvaluation #LLMBenchmarking #AppSec