Technology

Local-First Architecture for Comprehensive AI Security

🔐

Local Execution Protocol

Your model's weights, architecture, and training data never leave your infrastructure. All tests run locally. For Vision testing, the encrypted result includes each image's top-5 predictions and image hashes. For LLM, RAG, and Agent testing, the encrypted result includes the test probes and brief response excerpts needed to substantiate your report.

How it works

1

Install

pip install rednblue
2

Authenticate

Secure token

3

Local Test

On your machine

4

Report

PDF + Certificate

Testing Domains

Computer Vision

  • Gradient-based perturbation attacks
  • Optimization-based attacks
  • Detection evasion (YOLO)
  • Gradient masking diagnostic

PyTorch / ONNX classifiers, YOLO (Ultralytics)

Language Models (LLM)

  • Jailbreak testing
  • System prompt extraction
  • Indirect injection
  • Token manipulation

OpenAI, Anthropic, Gemini, Mistral, Cohere

RAG Pipelines

  • Corpus poisoning
  • Context stuffing
  • Boundary probing
  • RAG injection

LangChain, LlamaIndex, ChromaDB, Pinecone

AI Agents

  • Tool hijacking
  • Goal drift
  • Privilege escalation

AutoGen, CrewAI, LangChain Agents

Backdoor Detection

  • Consistency testing (STRIP)
  • Trigger enumeration
  • Sensitivity probing

IP Protection

  • Extraction feasibility
  • Model fingerprinting
  • Consistency analysis

Audio & Speech Models Roadmap

  • Adversarial audio perturbations
  • Transcription attacks (ASR)
  • Voice vulnerability detection
  • Ambient noise robustness

Whisper, Wav2Vec2, SpeechBrain, custom ASR pipelines (in development, not yet active)

Explainability & Tooling

Explainability (XAI)

Each vision test can produce an attention map showing where the model looks, before and after the attack. Generated 100% locally.

  • Grad-CAM, Grad-CAM++ (classifiers)
  • EigenCAM (YOLO detectors)
  • Robust Integrated Gradients (resistant to explanation manipulation)
  • 3-panel view: original, adversarial, heatmap

Desktop app & CLI

A local desktop application alongside the CLI: model picker, live CPU/GPU/RAM monitoring, robustness radar and XAI panels.

  • Local model or HuggingFace import
  • Vision and LLM modes
  • Gold / Silver / Bronze / Fail / Incomplete grades

Security & Compliance

Ed25519

Signed certificates

EU Hosting

Scaleway, France

RGPD

Data minimisation

EU AI Act

Supports Article 15 files

Documentation