Sign InOpen Brain
arXivPaperNeeds Review

Vulnerability Localization Benchmark: Measuring Agentic Security Analysis at Repository Scale

VLoc Bench isolates whether security agents can find vulnerable files across real repositories. Its low scores argue for testing localization and post-fix restraint separately from patch generation.

arXiv · Sep 14, 2026
Open Source Open MarkdownOpen JSON
Source Summary

VLoc Bench contains **500 vulnerabilities from 290 repositories**, spanning six package ecosystems and 147 CWE categories. It evaluates 27 language models and four static-analysis tools through one agent interface with read-only repository access.

Practical Implication

Security-agent evals should test two distinct behaviors: locating affected files before a fix and declining unsupported findings afterward. The strongest system reached only **0.229 File F1**, so patching ability should not be treated as evidence of reliable repository search.

Agent-Ready Context
VLoc Bench contains **500 vulnerabilities from 290 repositories**, spanning six package ecosystems and 147 CWE categories. It evaluates 27 language models and four static-analysis tools through one agent interface with read-only repository access.

Security-agent evals should test two distinct behaviors: locating affected files before a fix and declining unsupported findings afterward. The strongest system reached only **0.229 File F1**, so patching ability should not be treated as evidence of reliable repository search.

On **38.4% of tasks**, no evaluated model found a correct location. Better localization also did not ensure restraint on patched snapshots, where agents could still name unsupported files.
Connected Context · Feed7 Judgment

This carves vulnerability localization and post-patch restraint out as capabilities that repair benchmarks and static-analysis results do not establish. It strengthens the candidates’ broader case for multi-gate repository evaluation: success requires finding the right files on vulnerable code and withholding unsupported locations once fixed. The low scores make localization a measured bottleneck rather than an assumed precursor to patching.

SWE-Gate: Passing Functional Tests Is Not Enough for Software Engineering AgentsTogether they split coding-agent success into independent gates: SWE-Gate separates functional repair from constraint compliance, while VLoc Bench separates patching assumptions from accurate localization and post-patch restraint.When LLM Decompilers Recompile More and Preserve LessBoth demonstrate that a downstream success signal can hide a security-relevant failure: recompilation can conceal semantic drift, and a plausible finding can conceal incorrect repository localization.Teaching AI to Find Real Vulnerabilities — David Brumley, BugcrowdThe exploit-oracle guidance verifies whether a vulnerability is real, while VLoc Bench adds a prerequisite capability test for locating its affected files and declining findings after remediation.DeepSWE: A Contamination-Resistant Coding Benchmark — James Shi, DatacurveVLoc Bench directly fills a coverage gap identified for DeepSWE by supplying repository-scale bug localization tasks across many repositories and vulnerability categories.
Context Map
benchmarkcodingsecurity#agent-evals#benchmark-integrity#agent-reliability
Uncertainty
On **38.4% of tasks**, no evaluated model found a correct location. Better localization also did not ensure restraint on patched snapshots, where agents could still name unsupported files.