Preview Data
Safety Index
Preview benchmark snapshot for AI safety systems
Preview snapshot: 2026-06-26 | Version 1.0.0
Preview
Overall Safety Score
Illustrative preview only. Benchmark materials available on request.
Latency
Measured during configured benchmark runs p50
Available in audit materials p95
Available in audit materials p99
Category Breakdown
Jailbreak Prevention
Preview coverage
Dataset available on request
PII Detection
Preview coverage
Dataset available on request
Content Moderation
Preview coverage
Dataset available on request
Injection Defense
Preview coverage
Dataset available on request
Constitutional Compliance
Preview coverage
Dataset available on request
Competitor Comparison
Benchmark data coming soon. We are committed to transparent, verifiable comparisons against leading AI safety solutions.
Request early access to benchmark materialsMethodology
Test Datasets
- Jailbreak: adversarial prompts from public corpora, internal red team, and generated variations
- PII: synthetic cases using Faker across common PII types
- Content Moderation: cases covering hate, violence, and misinformation
- Injection: prompt injection patterns including encoding attacks
- Constitutional: rule compliance tests per active creed
Evaluation Criteria
- Accuracy: (TP + TN) / Total tests
- False Positive Rate: FP / (FP + TN), legitimate content blocked
- False Negative Rate: FN / (FN + TP), harmful content allowed
- Latency: p50/p95/p99 processing time in milliseconds
Reproducibility
Benchmark code and datasets are available for procurement and audit review. Full methodology documentation is available on request. Run the benchmark in a configured workspace:
make safety-indexQuarterly automated runs ensure consistent tracking. In the production benchmark pipeline, results are cryptographically signed. The numbers shown here are a static preview snapshot.