Security evaluation
Compares code from different models and coding agents under a consistent set of real CVE scenarios.
github.com/Tencent/AICGSecEval · RESEARCH
CodeGuard
Uses 132 real CVE scenarios and executable Docker PoCs to produce reproducible security scores for AI-generated code and detect known vulnerability patterns reintroduced by code changes.
Core capabilities
COREDynamic validation
Runs executable PoCs and functional tests instead of relying only on static vulnerability features.
Regression detection
Checks pull requests through GitHub webhooks and returns reviewable evidence with the result.
Who it is for
Useful for researchers comparing AI-generated code under consistent CVE scenarios, executable PoCs, and functional tests.
Read the boundary first
This page describes research based on upstream AICGSecEval. NGJOO does not expose a hosted scanner here, and results cannot prove safety in untested cases.
Ways to use it
- Compare models or coding agents on the same CVE tasks rather than selecting only successful examples.
- Pair static signals with executable PoCs and functional tests to separate security and functionality failures.
- Check whether a code change reintroduces a known pattern and return triggering evidence to the reviewer.
How to review the result
- Run PoCs only in isolation and never aim exploit steps at an unknown production system.
- Passing known tasks does not establish the absence of other vulnerabilities; state coverage and negative results.
- Record versions of the upstream dataset, containers, and scoring scripts with every reproducibility claim.
Current status and evidence
2026-09-23The upstream Tencent/AICGSecEval repository was reachable on 2026-09-23. This site exposes no public CodeGuard scanning endpoint.
This page links to the public upstream repository. NGJOO has not run an independent benchmark, so upstream results are not presented as NGJOO test results.