Documentation

Table of Contents

1. Install the CLI:

   pip install rednblue

2. Create an account: on dashboard.rednblue.io/register.
3. Purchase a token: in the dashboard, go to Credits → Buy and choose the tier that matches your needs.
4. Export your token:

   # Linux / macOS
   export RNB_TOKEN=RB-XXXXXX-XXXXXX
   
   # Windows
   set RNB_TOKEN=RB-XXXXXX-XXXXXX

5. Run your first test:

   # Image classifierrnb preview --model resnet50.pth \
       --input ./images/ \
       --submit

6. Download your report: in the dashboard, under Tests, download the PDF report and the certificate.

• Python 3.10+ (recommended 3.11 or 3.12)
• pip 23+
• RAM: 8 GB minimum, 16 GB recommended
• GPU: optional but strongly recommended for vision (CUDA 11.8+ or MPS on Apple Silicon)
• Disk space: 2 GB
• Internet connection: required only for results submission (the audit itself runs 100% locally)

Standard installation in a single command:

pip install rednblue

Recommended: use a virtual environment.

python3 -m venv rednblue-env
source rednblue-env/bin/activate
pip install rednblue
rnb --version

Note for Windows

Some optional dependencies (desktop UI) require Microsoft C++ Build Tools on Windows. To avoid this, install the CLI only:

# CLI only (recommended)pip install rednblue

# With desktop UI# requires Microsoft C++ Build Toolspip install rednblue[ui]

If you see an error "Microsoft Visual C++ 14.0 or greater is required", see the Troubleshooting.

rnb --version
# Expected output : rnb 3.0.0

rnb --help
# Lists available commands

RednBlue auto-detects your GPU. To use CUDA, install PyTorch with the right driver:

# CUDA 11.8
pip install torch --index-url https://download.pytorch.org/whl/cu118

# CUDA 12.1
pip install torch --index-url https://download.pytorch.org/whl/cu121

# Verify availabilitypython -c "import torch; print(torch.cuda.is_available())"

To force the device: --device auto | cpu | cuda | mps.

Supported architectures: ResNet, VGG, DenseNet, MobileNet, EfficientNet, ConvNeXt, SqueezeNet, ShuffleNet, AlexNet, GoogLeNet, ViT, timm.

# Freelancer auditrnb preview --model resnet50.pth --input ./images/ --submit

# Enterprise audit (full suite)rnb preview --model resnet50.pth --input ./images/ --submit --tier enterprise

Assessed Dimensions

• Noise Resilience — stability against sensor noise and compression artefacts
• Input Perturbation Defense — resistance to subtle, calculated modifications
• Classification Consistency — stability under gradient-based stress
• Targeted Evasion Defense — resistance to deliberate misclassification attempts
• Black-Box Resilience — resistance without access to model internals
• Iterative Stress Tolerance — defense against sustained multi-step pressure
• Query-Limited Defense — resistance with restricted interaction budget
• Decision Boundary Stability — decision boundary robustness

The Freelancer tier assesses a subset of dimensions; the Enterprise tier assesses them all at multiple intensity levels.

Supported versions: YOLOv5, YOLOv8, YOLOv9, YOLOv10, YOLOv11 (via Ultralytics API).

# Freelancer
rnb preview --model-type yolo --model best.pt --input ./images/ --submit

# Enterprise
rnb preview --model-type yolo --model best.pt --input ./images/ --submit \
    --tier enterprise --device cuda

Assessed Dimensions

• Noise Resilience — stability against environmental interference
• Input Perturbation Defense — resistance to imperceptible modifications
• Detection Consistency — detection reliability under stress
• Targeted Evasion Defense — resistance to forced disappearances/appearances
• Multi-Object Stability — accuracy on crowded scenes
• Object Persistence — detection persistence under perturbation
• Black-Box Resilience — resistance without model access
• Iterative Stress Tolerance — multi-step resistance
• Query-Limited Defense — query-efficient defense

Built-in gradient masking sanity check (consistency verification between gradient-based and black-box methods).

RednBlue parses your Python code, identifies the system type (plain LLM, RAG, autonomous agent), and recommends the matching assessment suite.

# Auto-detect + full auditrnb llm --file my_chatbot.py --auto

# Direct API providerrnb llm --provider openai --model gpt-4 --api-key $OPENAI_KEY

# Custom HTTP endpointrnb llm --endpoint https://api.example.com/chat \
    --auth-header "Authorization: Bearer $TOKEN"

Supported providers: OpenAI, Anthropic, Gemini, Mistral, Cohere, and any compatible HTTP endpoint.
Recognised frameworks: major chaining and vector storage frameworks.

Multiple dimension families, each tested across several scenarios to cover different levels of sophistication.

• Guardrail Resistance — maintenance of safety guardrails against adversarial queries
• System Prompt Confidentiality — protection against internal instruction leakage
• Indirect Injection Defense — resistance to injections via external content
• Tokenisation Robustness — stability against character-level manipulations
• RAG Corpus Integrity — detection of poisoning and contextual manipulation
• Agent Tool Safety — resistance to tool hijacking and privilege escalation
• Goal Stability — maintenance of the initial objective across long executions
• Backdoor Detection — identification of trigger-activated behaviours
• Extraction Resistance — protection against model reconstruction attempts

rnb preview

Runs a vision robustness audit (classifier or detector).

rnb preview [OPTIONS]

--model PATH              Path to model file (.pt, .pth, .onnx)--model-type TYPE         classifier (default) | yolo
--input PATH              Test image folder--tier TIER               freelancer (default) | enterprise
--device DEVICE           auto (default) | cpu | cuda | mps
--submit                  Submit results to the dashboard--verbose                 Verbose output

rnb llm

Runs an LLM, RAG, or agent audit.

rnb llm [OPTIONS]

--file PATH               Python file containing the pipeline--provider NAME           openai | anthropic | gemini | mistral | cohere
--model NAME              Model name--api-key STRING          API key (or via env variable)--endpoint URL            Custom HTTP endpoint--auto                    Auto-detect the suite--tier TIER               freelancer | enterprise

rnb status

Shows token and active session status.

rnb status              # current token statusrnb status --sessions   # list sessions

rnb ui

Launches the desktop UI (requires pip install rednblue[ui]).

rnb ui

Each token grants:

• 3 CLI runs — to iterate on settings
• 1 PDF report download
• 1 certificate download

The token is consumed at report download. The certificate can be re-downloaded while the token is valid.

"Microsoft Visual C++ 14.0 or greater is required" (Windows)

Cause: an optional dependency (desktop UI) requires a C++ compiler on Windows.

Solution:

# Option 1: install CLI only (recommended)pip install rednblue

# Option 2: install the Build Tools# https://visualstudio.microsoft.com/visual-cpp-build-tools/
# then:pip install rednblue[ui]

"CUDA not available" / GPU not detected

# Verify CUDApython -c "import torch; print(torch.cuda.is_available(), torch.version.cuda)"

# Reinstall PyTorch with CUDApip uninstall torch torchvision -y
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

# Force CPU if no GPUrnb preview --device cpu ...

"Token validation failed" / "Token consumed"

Possible causes:

• Token expired (1 year since purchase)
• Token already consumed (report downloaded)
• RNB_TOKEN variable incorrectly set

rnb status

export RNB_TOKEN=RB-XXXXXX-XXXXXX  # Linux/macOS
set RNB_TOKEN=RB-XXXXXX-XXXXXX     # Windows

"Submit failed" / network error

# Connectivity testcurl -I https://api.rednblue.io/health

# If behind a corporate proxyexport HTTPS_PROXY=http://proxy.company.com:8080
rnb preview ... --submit

"CUDA out of memory"

# Reduce batch sizernb preview ... --batch-size 1

# Force CPUrnb preview ... --device cpu

"Unsupported model architecture"

Verify your model is in the supported architectures list. For Hugging Face models, avoid timm/ repositories (not directly loadable as classifiers).