Sovereign · Self-improving · EU-native

I build AI security systems that defend themselves — and prove it with numbers.

Two-layer LLM defense (execution-time policy gate + input-time injection defense), measured against versioned adversarial corpora. Fully air-gappable, CPU-only, reproducible by anyone.

See the benchmark → What is Tharven?

100%

clear-attack block-rate

false positives

91.7%

obfuscation-bypass*

0.67ms

p95 latency

* The one we publish on purpose. A deterministic gate is near-blind to obfuscation — that honest number is the evidence that justifies a second, semantic layer. Production policy engine, measured 2026-06-04 on a SHA-pinned corpus. Full methodology →

Why Tharven is different

Most "AI security" projects ship zero reproducible numbers. The differentiator is honest, auditable measurement.

Defense-in-depth two layers

Execution-time policy gate blocks dangerous actions before they run; input-time defense blocks prompt-injection before it reaches the model. Measured separately, honestly.

Sovereign by design air-gapped

Runs fully offline, CPU-only, no third-party calls. Built for EU regulated environments where data cannot leave the premises — a structural advantage cloud-native incumbents can't match.

Reproducible methodology open

A published corpus + harness anyone can run in one command. A number always comes with its corpus SHA-256 and date. Weaknesses are published, not hidden.

Built for the AI Act Art. 15

EU AI Act Article 15 requires declared robustness and cybersecurity levels for high-risk AI. A declared number needs a reproducible test behind it. Tharven is that test.