Framedby Reference AI Security Labs
OPEN-SOURCE AI SECURITY FRAMEWORK

Read the research.
Run the experiment.

Framedby publishes free AI security engineering research paired with reproducible local experiments. Audit model weights, benchmark prompt injection defenses, and stress-test LLM deployments locally.

Browse Publications Launch Local Labs
100%
Reproducible
45+
Red Team Labs
0KB
Telemetry Leak

LOCAL LAB / VERIFIED PYTHON RUN

Agent Authority Boundary Lab

Example command after installing the bundled framework

python -m framedby_lab verify examples/authority-boundaries
Run locally
Engineering Pillars

Designed for Security Researchers & AI Architects

Traditional application security tools fall short when auditing stochastic neural networks. Framedby bridges the gap between theoretical research and local execution.

Reproducible Local Experiments

Run the first lab with synthetic inputs in local Python. Its bundle also includes an optional Docker Compose configuration.

  • Versioned lab bundles and checksums
  • Offline local HuggingFace cache

Threat & Vulnerability Intel

Continuous updates on prompt injection vectors, jailbreak taxonomies, model weights backdoors, and data extraction vulnerabilities.

  • OWASP Top 10 for LLMs coverage
  • Automated payload generation

Open Technical Papers

Peer-reviewed research whitepapers detailing deep mathematical and structural analysis of neural network security boundaries.

  • Free & open access PDFs
  • Complete code & dataset proofs
Interactive Catalog

Publications & Hands-on Labs

Verified · v1.0.0

Agent Authority Boundary Lab

Trace an allowed read, a denied write, and an authorized replay through a deterministic tool boundary.

Authorization · Red Teaming

Run locally →Publication

Planned

Bypassing Guardrails via Indirect Multi-Turn Injection

Research concept in the roadmap.

Prompt Injection

Planned

Steganographic Backdoors in SafeTensors Weights

Research concept in the roadmap.

Weights Auditing

Planned

Automated Adversarial Token Optimization (AATO)

Research concept in the roadmap.

Red Teaming

Planned

System Prompt Extraction via Model Calibration Exploits

Research concept in the roadmap.

Prompt Injection

Planned

Zero-Trust Proxy Gateway for Inference Endpoints

Research concept in the roadmap.

Red Teaming

Planned

Training Time Poisoning Mitigation Framework

Research concept in the roadmap.

Weights Auditing

Part of NeuralStack | MS - AI Security Engineering Ecosystem

Empowering AI Engineers with Transparent Defenses

Framedby is an open project hosted by NeuralStack | MS AI Security Engineering Ecosystem. We believe AI systems cannot be secured behind proprietary black boxes. By providing open-source, reproducible local setups, engineers can audit and fortify models with total sovereignty.

Contribute on GitHub Read Documentation