KNOWLEDGE BASE • SPECIFICATIONS • PAPERS

Resources

Technical documentation, engineering dispatches, and public security architecture whitepapers for agentic AI defense.

DEVELOPER SPECIFICATION

Documentation

API reference (cURL) and our measured pipeline benchmarks. SDK integration guides and MCP client setup: coming soon.

Key Topics Covered:

•5-Minute Quickstart & Authentication

•Transparent Proxy: /v1/chat/completions

•Standalone Inbound & Tool Result Scanners

•MCP Server stdio Supervisor Protocol (coming soon)

SECURITY & GOVERNANCE

Security Whitepapers

Technical whitepapers detailing EU-hosted CPU-only compute nodes, MITRE ATLAS taxonomy, and honest public limitations.

Key Topics Covered:

•Zero-Trust Sovereign Infrastructure Whitepaper

•Memory Vector Topology & MINJA Defense Paper

•Public Architectural Limitations & Boundaries

•EU AI Act Article 12 Compliance Specification

EMPIRICAL MEASUREMENT & BENCHMARKS

Our evaluations.

3 THIRD-PARTY TEST SUITES

Evaluations we've run on our own production system against third-party corpora and frameworks. Methodology, datasets, and aggregate results. This is not research — it is measurement.

EVALUATION · THIRD-PARTY CORPUSOCTOBER 2026

safelabs-eval: OWASP ASI Top 10 Evaluation

Methodology:

300 adversarial prompts across 10 OWASP ASI categories. Third-party corpus, scored by the library's own pattern detectors.

Result:0 VULNERABLE · 0 FAIL

What We Learned:

The corpus exposed three classes we had normalized past. All three are now covered at the class level.

EVALUATION · FRAMEWORK MAPPINGOCTOBER 2026

AMTSO mapping

Methodology:

Mapping of our tests to AMTSO vectors (not a certification). Five vectors mapped to detection layers with MITRE ATLAS technique IDs.

Result:5/5 vectors mapped · 10/11 end-to-end PASS (self-run)

What We Learned:

One miss was a test-calibration artifact, documented honestly. The underlying detector passes at 10/10 on its dedicated suite.

// We do not publish detection logic, specific patterns, or thresholds. We publish what we measured and what we learned.

DIRECT INTEGRATION ASSETS

Start testing in seconds

Everything needed to verify agent trust boundaries and evaluate zero-trust rules against your production model upstreams.

Ready to integrate the AI Firewall?

Request early access — the Community tier includes 5,000 monthly inspections free.