model-security

Toolkit to assess and determine model provenance

Veil Armor is an enterprise-grade security framework for Large Language Models (LLMs) that provides multi-layered protection against prompt injections, jailbreaks, PII leakage, and sophisticated attack vectors.

Security research on AI/ML model vulnerabilities based on DEF CON 33 presentations. Demonstrates pickle RCE, TorchScript exploitation, ONNX injection, model poisoning, and integrated LLM attacks with PromptMap2.

Educational research demonstrating weight manipulation attacks in SafeTensors models. Proves format validation alone is insufficient for AI model security.

LLM Sentinel Red Teaming Platform is an enterprise-grade framework for automated security testing of Large Language Models, detecting vulnerabilities such as jailbreaks, prompt injection, and system prompt leakage across multiple providers, with structured attack orchestration, risk scoring, and security reporting to harden models before production

Collection of Python security analysis tools for ML models and infrastructure. Includes FGSM harness, model inspection, poison monitoring, and deployment security validation.

GitHub Actions CI/CD pipeline for automated AI model security scanning with Palo Alto Networks Prisma AIRS

🛡️ Open-source AI security scanner & LLM red-teaming platform. Test LLM APIs, chatbots, agents, MCP servers & RAG for prompt injection, jailbreaks, data leaks & unsafe tool use — with OWASP LLM Top 10 mapping and plain-English, audit-ready reports.

🥒 Educational PyTorch pickle deserialization RCE demo showing how malicious .pt model files can execute arbitrary code during unsafe loading with torch.load().

Static scanner that detects code-execution backdoors in PyTorch/pickle ML model files (pickle-deserialization RCE), with an offensive demo generator. Python, stdlib-only.

Cryptographic provenance verification and binary inspection for ML model artifacts (Safetensors, GGUF, PyTorch) in CI/CD pipelines. Companion toolkit to the Help Net Security column Weaponized Weights.

Static security scanner for LoRA adapters (.safetensors) — M1 static analyzer for weight-level anomalies.

ML-infrastructure-aware anomaly detection system for protecting model weights against exfiltration, using a 3-layer cascaded architecture (Rules → ML → LLM).

AI Evaluator Pro 🛡️ is an AI security auditing tool that checks Hugging Face models for supply chain risks, unsafe formats, and author trust using OSINT + LLMs. It supports direct or discovery-based audits to detect security and integrity issues before deployment.

Six attack-and-defence projects against ML systems on one harness — gradient inversion, model extraction, pickle scanning, weight poisoning, adversarial detection and a prompt-injection firewall — reporting through one finding schema.

Lab: detect malicious AI models in Azure ML with Microsoft Defender for Cloud. Pickle RCE, exposed secrets, and unsafe opcodes caught before production. Includes Bicep, Sentinel analytics, workbook, and GitHub Actions SARIF gate.

Adversarial testing methodology, attack simulations, and MITRE ATLAS mappings for securing clinical AI and LLM systems.

MITRE ATT&CK-grounded security fine-tuning dataset (24,652 records, 18 sources, 100% per-record provenance) + extraction pipeline (23 upstream-source converters) + privacy-audit harness (Carlini 2021 prefix-completion extraction, Carlini 2022 MIA, MUSE 2024 per-token, LiRA). 106+ hermetic tests. GH-only distribution; consumed by Veedubin/AttackLM.

First-line security scanner for local AI model weights before AI-dept handoff. Deep ONNX protobuf parsing + pickle/safetensors/GGUF static analysis, HF provenance & SHA256 checks for vLLM serving. Stdlib-only core; optional [onnx] extra.