intertwine-ai/security-verifiers-e2-metadata

🔒 Security Verifiers E2: Security Configuration Verification (Public Metadata)

⚠️ This is a PUBLIC metadata-only repository. The full datasets are hosted privately to prevent training contamination. See below for access instructions.

Overview

E2 is a tool-grounded configuration auditing environment for Kubernetes and Terraform. This repository contains only the sampling metadata that describes how the private datasets were constructed.

Why Private Datasets?

Training contamination is a critical concern for benchmark integrity. If datasets leak into public training corpora:

• Models can memorize answers instead of learning to reason
• Evaluation metrics become unreliable
• Research reproducibility suffers
• True capabilities become obscured

By keeping evaluation datasets private with gated access, we:

• ✅ Preserve benchmark validity over time
• ✅ Enable fair model comparisons
• ✅ Maintain research integrity
• ✅ Allow controlled access for legitimate research

Dataset Composition

The private E2 datasets include:

Kubernetes Configurations

• Source: Real-world K8s manifests from popular open-source projects
• Scans: KubeLinter, Semgrep, OPA/Rego policies
• Violations: Security misconfigurations, best practice violations
• Severity: Categorized (high/medium/low) based on tool outputs

Terraform Configurations

• Source: Infrastructure-as-code from real projects
• Scans: Semgrep, OPA/Rego policies, custom rules
• Violations: Security risks, compliance issues
• Severity: Weighted scoring for reward computation

What's in This Repository?

This public repository contains:

1. Sampling Metadata (sampling-*.json):

   • Source repository information
   • File selection criteria
   • Scan configurations
   • Label distributions
   • Reproducibility parameters

2. Tools Versions (tools-versions.json):

   • KubeLinter version (pinned)
   • Semgrep version (pinned)
   • OPA version (pinned)
   • Ensures reproducible scanning

3. This README: Instructions for requesting access

Reward Components

E2 uses tool-grounded reward functions:

• Detection Precision/Recall/F1: Against ground-truth violations
• Severity Weighting: Higher reward for catching critical issues
• Patch Delta: Reward for proposed fixes that eliminate violations
• Re-scan Verification: Patches must pass tool validation

Multi-turn performance: Models achieve ~0.93 reward with tool calling vs ~0.62 without tools.

Requesting Access

🔑 To access the full private datasets:

1. Open an access request issue: Security Verifiers Issues
2. Use the title: "Dataset Access Request: E2"
3. Include:
   • Your name and affiliation
   • Research purpose / use case
   • HuggingFace username
   • Commitment to not redistribute or publish the raw data

Approval criteria:

• Legitimate research or educational use
• Understanding of contamination concerns
• Agreement to usage terms

We typically respond within 2-3 business days.

Citation

If you use this environment or metadata in your research:

@misc{security-verifiers-2025,
  title={Open Security Verifiers: Composable RL Environments for AI Safety},
  author={intertwine},
  year={2025},
  url={https://github.com/intertwine/security-verifiers},
  note={E2: Security Configuration Verification}
}

Related Resources

• GitHub Repository: intertwine/security-verifiers
• Documentation: See EXECUTIVE_SUMMARY.md and PRD.md in the repo
• Framework: Built on Prime Intellect Verifiers
• Other Environments: E1 (Network Logs), E3-E6 (in development)

Tools

The following security tools are used for ground-truth generation:

• KubeLinter: Kubernetes YAML linting and security checks
• Semgrep: Pattern-based static analysis for K8s and Terraform
• OPA: Policy-as-code validation with Rego

License

MIT License - See repository for full terms.

Contact

• Issues: GitHub Issues
• Discussions: GitHub Discussions

---

Built with ❤️ for the AI safety research community