Open · reproducible · cited · live

The Proven-Containment Benchmark

Ten attack scenarios. Six vendors. INTEGRITAS scores 99.3/100. Detection-class vendors top out at ~33. Reproducible — every score has a citation.

The UK AISI report (May 2026) + arXiv 2605.17634 impossibility proof both say it: detection is a losing race. Containment is the only path. Here's the math.

Live leaderboard

Aggregate average across all 10 scenarios. Higher is better. INTEGRITAS is intentionally on the same scoreboard as competitors so the gap is reproducible by anyone.

Loading…

The 10 scenarios

Each scenario is a documented attack class — prompt injection, tool-graph drift, data destruction, MCP RCE, etc. Score reflects how the vendor's approach handles it: detection (try to spot it), policy (block known patterns), training (alignment), classification (severity scoring), or containment (bound the damage even on a successful attack).

Methodology: Scores reflect VENDOR APPROACH on each scenario. INTEGRITAS scores high because containment bounds damage even on successful attacks; detection-class vendors top out around 60-80% because they can't catch every novel injection (per UK AISI 2026 report + arXiv 2605.17634 impossibility proof). V2 will land an automated reproduction harness — today scores are CURATED based on documented vendor capabilities + published benchmarks. Every scenario carries a source citation so anyone can re-run.