VulnGym: A Real-World Project-Level Vulnerability Benchmark

A benchmark for white-box vulnerability-hunting agents designed to evaluate detection capabilities within real-world engineering contexts. It features verifiable vulnerability trigger paths and code-semantic evidence chains. The dataset was created by Tencent and was last updated on 2026-05-25.

Use Cases

• Benchmarking vulnerability-hunting agents based on real-world project-level code
• Evaluating detection capabilities based on verifiable vulnerability trigger paths
• Training AI agents for white-box security analysis based on code-semantic evidence chains

Strengths

• Designed for real-world project-level evaluation contexts
• Includes verifiable vulnerability trigger paths and code-semantic evidence chains

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment

Provenance

Source

Tencent

Freshness

Last updated 2026-05-25 04:31:30; freshness should be verified