Skip to content

Repository files navigation

SafetyGap 🛡️

Mapping Multilingual Blind Spots in AI Safety Evaluation Infrastructure

Track 4 Hackathon

The Problem

International AI Safety Institutes cannot coordinate on safety evaluation if they lack shared visibility into what evaluation infrastructure actually exists.

Our analysis reveals:

  • 75% of EU languages have no jailbreak resistance benchmarks
  • 96% of EU languages have no bias detection benchmarks
  • 96% of EU languages have no adversarial robustness benchmarks

A model could pass every available safety evaluation in Greek, Hungarian, or Polish while having completely untested vulnerabilities in those languages.

The Solution

SafetyGap is an open-source tool that:

  1. Maps coverage — Which benchmarks exist for which languages
  2. Identifies gaps — Which language × category combinations have zero coverage
  3. Prioritizes action — Which gaps affect the most people
  4. Enables coordination — Shareable format for international collaboration

Quick Start

# View the interactive dashboard
# Open SafetyGapDashboard.jsx in any React environment

# Or run the Python analysis
python analyze_coverage.py

Key Findings

Risk Category EU Language Coverage Status
Truthfulness 22/24 (92%) ✅ Adequate
Over-refusal 24/24 (100%) ✅ Adequate
Toxicity 14/24 (58%) ⚠️ Moderate gap
Jailbreak 2/24 (8%) 🚨 Critical
Bias 1/24 (4%) 🚨 Critical
Adversarial 1/24 (4%) 🚨 Critical

Who Should Use This

Stakeholder Use Case
🏛️ AI Safety Institutes Coordinate benchmark development, avoid duplication
⚖️ Regulators Understand which compliance claims are verifiable
🔬 Benchmark Developers Identify high-impact targets for new work

Files

  • SafetyGapDashboard.jsx — Interactive React visualization
  • benchmarks.json — Structured benchmark database
  • submission_report.md — Full hackathon report

Contributing

This tool is designed to be community-maintained. To add a benchmark:

  1. Fork this repo
  2. Add entry to benchmarks.json
  3. Submit PR

References

  • Deng et al. (2024). Multilingual jailbreak challenges in LLMs. ICLR 2024.
  • Wang et al. (2024). All languages matter: Multilingual safety of LLMs. ACL 2024.
  • Mazeika et al. (2024). HarmBench. arXiv:2402.04249.

License

MIT


Track 4: International Verification & Coordination — Technical AI Governance Hackathon 2026

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages