Awesome AI Security

A curated list of awesome AI security related frameworks, standards, learning resources and open source tools.

If you want to contribute, create a PR or contact me @ottosulin.

Table of Contents

• Learning Resources
  • Reading & Guides
  • Courses, Labs & CTFs
  • Podcasts
• Governance & Risk Management
  • Frameworks
  • Standards & Verification
  • Taxonomies, Terminology & Risk Databases
  • Checklists & Practical Guidance
• Attack Techniques & Red Teaming
  • Adversarial ML & Classical Models
  • LLM & GenAI Red Teaming
  • Agentic AI & MCP Attack Tools
  • AI-Assisted Offensive Security
  • Steganography & Covert Channels
• Benchmarks & Evaluations
• Defense & Security Controls
  • Input/Output Guardrails
  • Agent Runtime Security & Sandboxing
  • MCP Security
  • Model & Artifact Scanning
  • AI-Assisted Defensive Security
  • Privacy & Confidential Computing
  • Data & Supply Chain Security
• Agentic AI Security Skills
• Security-Focused AI Models

Learning Resources

Reading & Guides

• OWASP ML TOP 10
• OWASP LLM TOP 10
• OWASP AI Security and Privacy Guide
• NIST AIRC - NIST Trustworthy & Responsible AI Resource Center
• The MLSecOps Top 10 by Institute for Ethical AI & Machine Learning
• OWASP Multi-Agentic System Threat Modeling
• OWASP: CheatSheet – A Practical Guide for Securely Using Third-Party MCP Servers 1.0
• OWASP GenAI Threat & Defense Compass - Reference guide mapping GenAI threats to corresponding defensive controls.
• NCSC Guidelines for Secure AI System Development - Practical guidelines co-authored by NCSC (UK) and CISA for secure AI development, covering design, development, deployment, and operation.
• CoSAI – Preparing Defenders for a Changing Cybersecurity Landscape - Coalition for Secure AI workstream preparing defenders for AI-era threats.

Courses, Labs & CTFs

• Damn Vulnerable MCP Server - A deliberately vulnerable implementation of the Model Context Protocol (MCP) for educational purposes.
• otto-support - A vulnerable MCP server implementation using mcp-go with tiered authentication (4 roles), 19 tools, and built-in sandboxed container with Claude Code for practicing privilege escalation and tool misuse.
• OWASP WrongSecrets LLM exercise
• vulnerable-mcp-servers-lab - A collection of servers which are deliberately vulnerable to learn Pentesting MCP Servers.
• FinBot Agentic AI Capture The Flag (CTF) Application - FinBot is an Agentic Security Capture The Flag (CTF) interactive platform that simulates real-world vulnerabilities in agentic AI systems using a simulated Financial Services-focused application.
• AI-Red-Teaming-Playground-Labs - AI Red Teaming playground labs to run AI Red Teaming trainings including infrastructure.
• Damn Vulnerable LLM Agent - Intentionally vulnerable LLM agent for learning about prompt injection, tool misuse, and agent security
• • LLMVault - An open-source CTF-style LLM security lab for learning the OWASP LLM Top 10 through hands-on vulnerable exercises covering prompt injection, RAG attacks, system prompt leakage, tool misuse, and agent security.
• AIGoat - Open-source AI security playground for LLM red teaming: 17 attack labs covering the full OWASP LLM Top 10, 9 CTF challenges, and 3 progressive defense levels, running locally with Ollama.

Podcasts

• MLSecOps podcast
• AI Security Podcast
• AI Security Ops - Weekly podcasts from Black Hills Information Security exploring how AI transforms cybersecurity—covering emerging threats, tools, and trends with practical, actionable knowledge.
• GenAI Security podcast

Governance & Risk Management

Frameworks

• NIST AI Risk Management Framework
• ISO/IEC 42001 Artificial Intelligence Management System
• ISO/IEC 23894:2023 Information technology — Artificial intelligence — Guidance on risk management
• Google Secure AI Framework (SAIF)
• ENISA Multilayer Framework for Good Cybersecurity Practices for AI
• OWASP Artificial Intelligence Maturity Assessment
• CSA AI Model Risk Framework
• CSA Maestro AI Threat Modeling Framework
• CSA AI Controls Matrix - Comprehensive controls matrix for AI systems covering governance, risk, and compliance.
• OWASP Agentic AI Top 10 - Top 10 for Agentic AI serving as the core for OWASP and CSA red teaming.
• OWASP GenAI Crosswalk - Interactive crosswalk mapping OWASP GenAI risks (LLM Top 10, Agentic Top 10, DSGAI 2026) to 25 industry frameworks with coverage scoring and gap analysis.
• NIST SP 800-218A Secure Software Development Practices for Generative AI and Dual-Use Foundation Models - Extension of SSDF (SP 800-218) with specific practices for GenAI and dual-use foundation model development.
• NIST AI 600-1 Generative AI Profile - Companion profile to NIST AI RMF 1.0 with specific risk actions for generative AI systems including unique risks like hallucination, data poisoning, and privacy.

Standards & Verification

• OWASP AI Security Verification Standard
• OWASP Agent Name Service
• OWASP Agent Observability Standard
• AI Verify - Singapore government-backed AI testing framework and toolkit for verifying AI system properties against governance frameworks.
• Agent Control Standard - Open standard for runtime agent control: declarative hooks, policy enforcement, and observability across AI agent frameworks (OWASP GenAI Security Project).
• AARM - Cloud Security Alliance specification defining the system category and capabilities for agentic runtime security.

Taxonomies, Terminology & Risk Databases

• NIST AI 100-2e2023 - Adversarial machine learning taxonomy and terminology
• MITRE ATLAS
  • ATLAS Knowledge Base Agent - Open-source AI assistant for exploring the ATLAS knowledge base with natural language, MCP server access, and customizable agent workflows.
• AVIDML
• MIT AI Risk Repository
• AI Incident Database
• ISO/IEC 22989:2022 Information technology — Artificial intelligence — Artificial intelligence concepts and terminology
• NIST AI Glossary
• The Arcanum Prompt Injection Taxonomy - Comprehensive prompt injection attack classification system covering attack intents, techniques, evasions, and input vectors. Categorizes goals, methods, obfuscation techniques, and attack surfaces for prompt injection attacks.
• CSA LLM Threats Taxonomy
• OWASP AI Vulnerability Scoring System (AIVSS) - Scoring system for AI-specific vulnerabilities, extending CVSS concepts to AI risk dimensions.

Checklists & Practical Guidance

• OWASP LLM Applications Cybersecurity and Governance Checklist
• OWASP LLM and Generative AI Security Center of Excellence Guide
• OWASP Agentic AI – Threats and Mitigations
• OWASP AI Security Solutions Landscape
• OWASP GenAI Incident Response Guide
• OWASP LLM and GenAI Data Security Best Practices
• OWASP Securing Agentic AI Applications
• OWASP GenAI Red Teaming Guide
• CSA Secure LLM Systems: Authorization Practices - Essential authorization practices for LLM-backed systems.

Attack Techniques & Red Teaming

Adversarial ML & Classical Models

• Adversarial Robustness Toolkit - ART focuses on the threats of Evasion (change the model behavior with input modifications), Poisoning (control a model with training data modifications), Extraction (steal a model through queries) and Inference (attack the privacy of the training data)
• cleverhans - An adversarial example library for constructing attacks, building defenses, and benchmarking both
• foolbox - A Python toolbox to create adversarial examples that fool neural networks in PyTorch, TensorFlow, and JAX
• TextAttack - A Python framework for adversarial attacks, data augmentation, and model training in NLP
• Counterfit - generic automation layer for assessing the security of machine learning systems
• OffsecML Playbook - A collection of offensive and adversarial TTPs with proofs of concept
• BadDiffusion - Official repo to reproduce the paper "How to Backdoor Diffusion Models?" published at CVPR 2023
• secml-torch - SecML-Torch: A Library for Robustness Evaluation of Deep Learning Models

LLM & GenAI Red Teaming

• garak - security probing tool for LLMs
• ai-scanner - Open-source web application for AI model security assessments, built on NVIDIA garak. Features 179 probes, multi-target scanning, scheduled scans, ASR scoring, and SIEM integration.
• promptfoo - Test your prompts, agents, and RAGs. Red teaming, pentesting, and vulnerability scanning for LLMs. Compare performance of GPT, Claude, Gemini, Llama, and more. Simple declarative configs with command line and CI/CD integration.
• PyRIT - The Python Risk Identification Tool for generative AI (PyRIT) is an open source framework built to empower security professionals and engineers to proactively identify risks in generative AI systems.
• PurpleLlama - Set of tools to assess and improve LLM security.
• augustus - LLM security testing framework for detecting prompt injection, jailbreaks, and adversarial attacks. 190+ probes, 28 providers, single Go binary. Production-ready with concurrent scanning, rate limiting, and retry logic.
• FuzzyAI - A powerful tool for automated LLM fuzzing. It is designed to help developers and security researchers identify and mitigate potential jailbreaks in their LLM APIs.
• EasyJailbreak - An easy-to-use Python framework to generate adversarial jailbreak prompts.
• gptfuzz - Official repo for GPTFUZZER: Red Teaming Large Language Models with Auto-Generated Jailbreak Prompts
• llamator - Framework for testing vulnerabilities of large language models (LLM).
• agentic_security - Agentic LLM Vulnerability Scanner / AI red teaming kit
• Agentic Radar - Open-source CLI security scanner for agentic workflows.
• RAPTOR - Autonomous offensive/defensive security research framework built on Claude Code. Chains static analysis (Semgrep, CodeQL), binary analysis, LLM-powered vulnerability validation, exploit generation, and patch writing. Multi-model orchestration with Z3-based feasibility analysis.
• whistleblower - Offensive security tool for testing against system prompt leakage and capability discovery of an AI application exposed through API
• promptmap - a prompt injection scanner for custom LLM applications
• spikee - Simple Prompt Injection Kit for Evaluation and Exploitation
• ps-fuzz - Make your GenAI Apps Safe & Secure — Test & harden your system prompt
• EasyEdit - Modify an LLM's ground truths
• llm-attacks - Universal and Transferable Attacks on Aligned Language Models
• llm-security - New ways of breaking app-integrated LLMs
• Plexiglass - A toolkit for detecting and protecting against vulnerabilities in Large Language Models (LLMs).
• Prompt Hacking Resources - A list of curated resources for people interested in AI Red Teaming, Jailbreaking, and Prompt Injection
• Giskard - Open-Source Evaluation & Testing for AI & LLM systems
• blackice - BlackIce is an open-source containerized toolkit designed for red teaming AI models, including Large Language Models (LLMs) and classical machine learning (ML) models. Inspired by the convenience and standardization of Kali Linux in traditional penetration testing, BlackIce simplifies AI security assessments by providing a reproducible container image preconfigured with specialized evaluation tools.
• ai-best-practices - Semgrep Pro Rules to ensure code using LLMs is following best practices. 58 rules, 102 sub-rules covering 6 providers + MCP + Claude Code & Cursor hooks + LangChain. Detects hardcoded API keys, prompt injection risks, missing safety checks, and unhandled errors across 7 languages.
• G0DM0D3 - Open-source multi-model chat interface for red teaming with 50+ models via OpenRouter. Features GODMODE CLASSIC (5 jailbreak combos), ULTRAPLINIAN multi-model evaluation, Parseltongue input perturbation engine with 33 red team techniques, and AutoTune adaptive sampling for AI safety research.
• L1B3RT4S - Collection of jailbreak and liberation prompts for major LLMs. Curated library of adversarial prompts designed to test and evaluate AI safety guardrails across ChatGPT, Claude, Gemini, and other frontier models.
• OBLITERATUS - Abliteration toolkit that removes refusal behaviors from open-source LLMs without retraining. Modifies model weights to eliminate safety-aligned refusal responses, enabling unrestricted model behavior research.
• T3MP3ST - Autonomous red teaming platform and multi-agent offensive-security meta-harness. Coordinates swarms of AI agents for coordinated adversarial testing of frontier AI systems.
• P4RS3LT0NGV3 - Universal text transformation and promptcrafting toolkit. Supports translation, mutation, encoding/decoding, and adversarial prompt engineering for jailbreak payload construction.
• claude-secure-coding-rules - Open-source security rules that guide Claude Code to generate secure code by default.
• DeepTeam - LLM red teaming framework with 40+ attack methods including prompt injection, jailbreaking, and RAG poisoning. Integrates with CI/CD pipelines.
• PoisonedRAG - Knowledge Corruption Attacks to Retrieval-Augmented Generation (USENIX Security 2025).

Agentic AI & MCP Attack Tools

• RAMPART - pytest-native safety and security testing framework for agentic AI applications.
• AI-Infra-Guard - A comprehensive, intelligent, and easy-to-use AI Red Teaming platform developed by Tencent Zhuque Lab. Integrates modules for Infra Scan, MCP Scan, and Jailbreak Evaluation, providing a one-click web UI, REST APIs, and Docker-based deployment for comprehensive AI security evaluation.
• OpenPromptInjection - A benchmark for prompt injection attacks and defenses
• AIMap - Internet-scale discovery and security testing platform for exposed AI agent infrastructure. Queries Shodan for MCP servers, Ollama instances, vLLM/LiteLLM proxies, and more — then fingerprints, scores risk, and launches protocol-specific attack suites with real-time streaming results.
• Humanbound - Open-source CLI that sends LLM-generated adversarial attacks to an agent's HTTP endpoint and scores the results against OWASP LLM Top 10, OWASP Agentic Top 10, NIST, and EU AI Act mappings.
• Agentic Chaos - Security chaos engineering for AI agents and LLM apps: injects reproducible faults (tool failures, prompt injection, MCP rug pulls, guardrail outages) and verifies security and resilience invariants still hold.

AI-Assisted Offensive Security

• PentestGPT - A GPT-empowered penetration testing tool

• HackingBuddyGPT - Helping Ethical Hackers use LLMs in 50 Lines of Code or less

• BugTraceAI - Open-source multi-agent platform for authorized web application security testing with validation, evidence capture, and reporting.

• cai - Cybersecurity AI (CAI), an open Bug Bounty-ready Artificial Intelligence (paper)

• shannon - Fully autonomous AI pentester for web apps and APIs by Keygraph. White-box security testing that analyzes source code, identifies attack vectors, and executes real exploits. 96.15% success rate (100/104 exploits) on XBOW benchmark.

• strix - Strix are autonomous AI agents that act just like real hackers - they run your code dynamically, find vulnerabilities, and validate them through actual proof-of-concepts

• redamon - AI-powered agentic red team framework that automates offensive security operations from reconnaissance to exploitation to post-exploitation with zero human intervention.

• CyberStrikeAI - AI-native security testing platform built in Go. Integrates 100+ security tools with an intelligent orchestration engine, role-based testing with predefined security roles, skills system, and comprehensive lifecycle management. Uses MCP protocol and AI agents for end-to-end automation from conversational commands to vulnerability discovery.

• HexStrikeAI - HexStrike AI MCP Agents is an advanced MCP server that lets AI agents (Claude, GPT, Copilot, etc.) autonomously run 150+ cybersecurity tools for automated pentesting, vulnerability discovery, bug bounty automation, and security research.

• mcp-for-security - A collection of Model Context Protocol servers for popular security tools like SQLMap, FFUF, NMAP, Masscan and more. Integrate security testing and penetration testing into AI workflows.

• mcp-security-hub - A growing collection of MCP servers bringing offensive security tools to AI assistants. Nmap, Ghidra, Nuclei, SQLMap, Hashcat and more.

• Burp MCP Server - MCP Server for Burp

• burpgpt - A Burp Suite extension that integrates OpenAI's GPT to perform an additional passive scan for discovering highly bespoke vulnerabilities and enables running traffic-based analysis of any type.

• guardian-cli - AI-Powered Security Testing & Vulnerability Scanner. Guardian CLI is an intelligent security testing tool that leverages AI to automate penetration testing, vulnerability assessment, and security auditing.

• AutoPentestX - AutoPentestX – Linux Automated Pentesting & Vulnerability Reporting Tool

• HackGPT - A tool using ChatGPT for hacking

• nano-analyzer - A minimal LLM-powered zero-day vulnerability scanner by AISLE.

• clearwing - Autonomous vulnerability scanner and source-code hunter built on LangGraph.

• Zen-AI-Pentest - AI-Powered Penetration Testing Framework with automated vulnerability scanning, multi-agent system, and compliance reporting. 72+ security tools, Docker sandbox, ReAct agents, attack path analysis.

• Violin - Supervised, agentic Hermes Agent pentest profile: 31 skill-based playbooks (OWASP Top 10, API Top 10, LLM Top 10) with interactive scoping, scope validation, and approval gates for authorised recon, exploit validation, and reporting.

• NeuroSploit - Autonomous multi-model pentesting harness in Rust. LLM pool drives recon, agent selection, exploit chaining, and cross-model validation voting across black-box, white-box, grey-box, and cloud engagement modes.

• OpenHack - AI-powered multi-agent scanner that autonomously finds SQLi, XSS, IDOR, and auth bypass in your codebase, then verifies each finding through sandbox execution and browser replay. On par with Claude Opus 4.6 at roughly 40x lower cost.

• PentAGI - Fully autonomous multi-agent system for complex penetration testing tasks. Sandboxed Docker execution, multi-provider LLM support (OpenAI, Anthropic, Gemini, Ollama, DeepSeek), knowledge graph integration, and real-time agent supervision.

• V3SP3R - AI-powered hardware hacking companion for the Flipper Zero. Natural language interface for controlling hardware attacks, with smart glasses integration for hands-free operation.

• DeepZero - MIT-licensed Windows driver research framework combining Ghidra decompilation, static analysis, and optional LLM assessment in resumable YAML pipelines.

Steganography & Covert Channels

• ST3GG - All-in-one steganography suite with multi-layer encoding, image and audio steganography, and steganalysis tools for detecting hidden data in AI-generated media.

Benchmarks & Evaluations

• ISC-Bench - Internal Safety Collapse: jailbreaks any frontier LLM (Claude Opus 4.6, GPT-5.4) in pass@3 via normal task completion — no adversarial prompting. Black-box, cross-domain, cross-science. Novel failure mode. [Paper]
• jailbreakbench - JailbreakBench: An Open Robustness Benchmark for Jailbreaking Language Models [NeurIPS 2024 Datasets and Benchmarks Track]
• AgentDojo - A Dynamic Environment to Evaluate Attacks and Defenses for LLM Agents.
• AIRTBench - Code Repository for: AIRTBench: Measuring Autonomous AI Red Teaming Capabilities in Language Models
• HackingBuddyGPT benchmark dataset - Benchmark dataset for automated pentesting
• AgentDoG - AgentDoG is a risk-aware evaluation and guarding framework for autonomous agents. It focuses on trajectory-level risk assessment, aiming to determine whether an agent's execution trajectory contains safety risks under diverse application scenarios.
• AICGSecEval - Tencent's comprehensive evaluation benchmark for AI code generation security, covering 10 CWE categories with automated test harness
• Inspect - Framework for large language model evaluations by the UK AI Security Institute. 200+ pre-built evaluations covering prompt engineering, tool usage, multi-turn dialog, and model-graded scoring.
• sec-code-bench - Alibaba's benchmark for evaluating LLM code security capabilities across 17 vulnerability categories and 4 programming languages
• BaxBench - Benchmark for evaluating LLM-generated backend code security — tests whether AI-generated web services contain exploitable vulnerabilities
• auto-pen-bench - Benchmark for evaluating automated penetration testing agents across network and web application scenarios

Defense & Security Controls

Input/Output Guardrails

• NeMo-Guardrails - NeMo Guardrails is an open-source toolkit for easily adding programmable guardrails to LLM-based conversational systems.
• LlamaFirewall - LlamaFirewall is a framework designed to detect and mitigate AI centric security risks, supporting multiple layers of inputs and outputs, such as typical LLM chat and more advanced multi-step agentic operations.
• llm-guard - LLM Guard by Protect AI is a comprehensive tool designed to fortify the security of Large Language Models (LLMs).
• Guardrails.ai - Guardrails is a Python package that lets a user add structure, type and quality guarantees to the outputs of large language models (LLMs)
• TrustGate - Generative Application Firewall (GAF) to detect, prevent and block attacks against GenAI Applications
• ZenGuard AI - The fastest Trust Layer for AI Agents
• LocalMod - Self-hosted content moderation API with prompt injection detection, toxicity filtering, PII detection, and NSFW classification. Runs 100% offline.
• DynaGuard - A Dynamic Guardrail Model With User-Defined Policies
• AprielGuard - 8B parameter safety–security safeguard model
• Safe Zone - Safe Zone is an open-source PII detection and guardrails engine that prevents sensitive data from leaking to LLMs and third-party APIs.
• superagent - Superagent provides purpose-trained guardrails that make AI-agents secure and compliant.
• ShellWard - AI Agent Security Middleware with 8-layer defense against prompt injection, data exfiltration & dangerous commands. Zero dependencies.
• CodeGate - An open-source, privacy-focused project that acts as a layer of security within a developer's Code Generation AI workflow
• Future AGI - Open-source self-hostable platform with built-in real-time guardrails for unsafe outputs (jailbreak, PII, injection, toxicity), evals, tracing, simulations, and gateway for LLM and agent applications.
• jes - Open-source guardrails for AI agents powered by decision models like Jev; checks prompts, retrieved content, tool calls and responses for prompt injection, jailbreaks and secret/PII leaks.
• Prompt Injection Defenses - Comprehensive collection of every practical and proposed defense against prompt injection.

Agent Runtime Security & Sandboxing

• OpenShell - OpenShell is the safe, private runtime for autonomous AI agents. It provides sandboxed execution environments governed by declarative YAML policies that prevent unauthorized file access, data exfiltration, and uncontrolled network activity.
• OpenSandbox - Secure, Fast, and Extensible Sandbox runtime for AI agents. Multi-language SDKs, Docker/Kubernetes runtimes, gVisor/Kata Containers/Firecracker isolation. CNCF Landscape project.
• CubeSandbox - Instant, concurrent, secure & lightweight sandbox for AI agents by Tencent Cloud. Sub-60ms cold start, <5MB memory overhead, E2B SDK compatible. Built on RustVMM and KVM with extreme isolation (dedicated kernel + eBPF).
• ADR - Enterprise agentic AI detection and response — observability, security benchmarking (300+ tasks, 133 MCP servers), and dual-agent threat detection. Deployed at Uber. MLSys 2026.
• Aegis - Open-source EDR for AI agents by Antropos. Monitor processes, files, network, and behavior of autonomous AI agents in real time. No telemetry, no cloud, everything stays local.
• Microsoft Agent Governance Toolkit - AI Agent Governance Toolkit from Microsoft — Policy enforcement, zero-trust identity, execution sandboxing, and reliability engineering for autonomous AI agents. Covers 10/10 OWASP Agentic Top 10.
• agentfield - Open-source control plane for agent systems with cryptographic identity, policy enforcement, and audit-friendly observability.
• leash - Leash wraps AI coding agents in containers and monitors their activity.
• Agent Safehouse - macOS sandbox for LLM coding agents using sandbox-exec with composable policy profiles and a deny-first model.
• vibekit - Run Claude Code, Gemini, Codex — or any coding agent — in a clean, isolated sandbox with sensitive data redaction and observability baked in.
• pipelock - Security harness for AI agents — egress proxy with DLP scanning, SSRF protection, MCP response scanning, and workspace integrity monitoring
• skill-scanner - A security scanner for AI Agent Skills that detects prompt injection, data exfiltration, and malicious code patterns. Combines pattern-based detection (YAML + YARA), LLM-as-a-judge, and behavioral dataflow analysis for comprehensive threat detection.
• SkillSpector - Security scanner for AI agent skills. Detects 64 vulnerability patterns across 16 categories with static analysis + optional LLM semantic evaluation. Multi-format output (JSON, Markdown, SARIF).
• Project CodeGuard - CoSAI Open Source Project for securing AI-assisted development workflows. CodeGuard provides security controls and guardrails for AI coding assistants to prevent vulnerabilities from being introduced during AI-generated code development.
• AgentLens - Agent observability and replay tooling for AI safety & interpretability research. Harness for running multi-session agent trajectories, capturing them in ATIF format, and tracking file state changes across sessions. Built for studying LLM agent behavior across multi-turn, multi-session, multi-agent interactions.
• OWASP Agent Memory Guard - Runtime defense layer for AI agent memory. Screens every memory read/write through detectors (prompt injection, secret/PII leakage, protected-key tampering, anomalies) and a YAML policy (allow/redact/quarantine/block), with SHA-256 integrity baselines, snapshots/rollback, and structured security events. Reference implementation for OWASP ASI06 Memory & Context Poisoning. Apache-2.0.
• claude-code-devcontainer - Sandboxed devcontainer for running Claude Code in bypass mode safely. Built for security audits and untrusted code review.
• coop - Rust CLI for running Claude Code and Codex inside disposable isolated VMs — Firecracker microVMs on Linux, Lima on macOS — with workspace sync, profiles, and commit/restore checkpoints.
• Quicksand - Async Python API to launch, control, and snapshot QEMU VMs with a focus on sandboxing AI agents. Pre-built Ubuntu/Alpine images, no root/Docker/system deps, works on macOS, Linux, and Windows. Includes checkpoint/revert and desktop control for agent sandboxes.
• claude-code-safety-net - A Claude Code plugin that acts as a safety net, catching destructive git and filesystem commands before they execute
• OneCLI - Open-source credential vault for AI agents. Rust HTTP gateway intercepts agent requests and injects API credentials transparently so agents never hold raw keys. AES-256-GCM encryption, per-agent scoping, full audit trail.
• openclaw-shield - Security plugin for OpenClaw agents - prevents secret leaks, PII exposure, and destructive command execution
• clawsec - Security scanner and hardening tool for OpenClaw deployments. Provides security assessments, configuration auditing, and vulnerability detection specifically for OpenClaw gateway and agent configurations.
• nanoclaw - Lightweight alternative to OpenClaw that runs in containers for security. Connects to WhatsApp, has memory, scheduled jobs, and runs directly on Anthropic's Agents SDK. First AI assistant to support Agent Swarms for collaborative agent teams.
• secureclaw - Automated security hardening for OpenClaw AI agents by Adversa AI. 51 audit checks, 12 behavioral rules, 9 scripts, 4 pattern databases. Full OWASP ASI Top 10 coverage. Protects against prompt injection, credential theft, supply chain attacks, and privacy leaks.
• defenseclaw - Enterprise governance layer for OpenClaw from Cisco AI Defense. Scans skills, MCP servers, and plugins with built-in CodeGuard SAST, tool call inspection engine, LLM guardrail proxy, and SIEM integration. Auto-blocks HIGH/CRITICAL findings.
• microsandbox - Lightweight microVM sandbox for running untrusted AI-generated code safely with strong isolation guarantees
• Adrian - Open-source, AARM-aligned runtime security monitoring and control engine for AI agents by Secure Agentics. Analyses agent activity logs (tool calls, actions, outputs) and reasoning traces to detect malicious, misaligned, or out-of-remit behaviour, with optional in-flight intervention (audit vs block mode). Python (LangChain/LangGraph) and TypeScript SDKs, fully self-hostable offline. Apache-2.0.
• HOL Guard - Local-first security harness that intercepts tool calls in AI coding agents (Codex, Claude Code, Cursor, Gemini, Copilot, Hermes, OpenCode) before files change or network is contacted. Pre-tool hooks, approval center, supply-chain advisory scanning, and optional Guard Cloud sync.
• Bernstein - Open-source governance layer for AI agents, covering CLI coding agents (Claude Code, Codex, Gemini CLI, Aider, and 40+ more) as well as self-hosted and hosted model endpoints. Runs each task in an isolated git worktree with per-agent credential scoping, PII gating, and a policy engine. No model sits in the coordination loop, so runs replay byte-identically, and an opt-in HMAC-chained audit log with signed per-artifact lineage can be verified offline with bernstein audit verify. Apache-2.0.
• Doberman - Runtime authorization layer between a coding agent and its tools. A local policy engine gives every tool call an allow/authenticate/block verdict before it executes; blocks carry reason codes, logs redact secrets to HMAC fingerprints, and errors fail closed. Ships an MCP proxy plus Claude Code and Codex adapters. Apache-2.0.
• Numbat - Endpoint visibility into AI agent activity from Perplexity. Observes desktop, CLI, IDE, and gateway agents via hooks, plugins, OTLP logs, and on-disk session artifacts; normalizes live and at-rest activity into one event model evaluated by a local CEL rule engine. Optional pre-action blocking and forensic reconstruction from session artifacts. Single Go binary, Apache-2.0.
• Substrate - Agent infrastructure runtime from Google that multiplexes stateful agent actors onto shared sandboxed workers (microVM, gVisor) with suspend/resume teleport, full-state snapshots, and kernel-level network isolation. Built on Kubernetes; hosts Google Agent Executor (ax) and kagent workloads. Apache-2.0.
• Prismor - Runtime security for AI coding agents and agent frameworks (Claude Code, Codex, Gemini CLI, Cursor, Copilot, LangChain, OpenAI Agents SDK and more). Pre-tool hooks check every tool call against a local YAML policy in observe or enforce mode, with secret cloaking, an MCP gateway, an LLM proxy for agents it cannot hook, egress control, and an Ed25519-signed audit trail. Apache-2.0.

MCP Security

• MCP-Security-Checklist - A comprehensive security checklist for MCP-based AI tools. Built by SlowMist to safeguard LLM plugin ecosystems.
• Awesome-MCP-Security - Everything you need to know about Model Context Protocol (MCP) security.
• secure-mcp-gateway - This Secure MCP Gateway is built with authentication, automatic tool discovery, caching, and guardrail enforcement.
• mcp-context-protector - context-protector is a security wrapper for MCP servers that addresses risks associated with running untrusted MCP servers, including line jumping, unexpected server configuration changes, and other prompt injection attacks
• mcp-guardian - MCP Guardian manages your LLM assistant's access to MCP servers, handing you realtime control of your LLM's activity.
• MCP Audit VSCode Extension - Audit and log all GitHub Copilot MCP tool calls in VSCode centrally with ease.
• MCP-Scan - A security scanning tool for MCP servers
• ATR (Agent Threat Rules) - Open-source detection rules for AI agent threats. 108 regex rules covering prompt injection, tool poisoning, credential exfiltration across 9 categories. Used by Cisco AI Defense. MIT licensed.
• MCPProxy - Local-first MCP proxy with per-tool SHA-256 quarantine to detect tool-poisoning and rug-pull attacks, automatic sensitive-data and secret scanning of tool calls, Docker sandbox isolation for untrusted MCP servers, and OAuth 2.1. MIT licensed.
• Gram - Open-source AI control plane that centrally manages MCPs, Skills, and Assistants with fine-grained permissions, real-time policy enforcement, threat detection, and event logs.

Model & Artifact Scanning

• modelscan - ModelScan is an open source project from Protect AI that scans models to determine if they contain unsafe code.
• picklescan - Security scanner detecting Python Pickle files performing suspicious actions
• fickling - A Python pickling decompiler and static analyzer
• medusa - AI-first security scanner with 74+ analyzers, 180+ AI agent security rules, intelligent false positive reduction. Supports all languages. CVE detection for React2Shell, mcp-remote RCE.
• julius - LLM service fingerprinting tool for security professionals. Detects 32+ AI services (Ollama, vLLM, LiteLLM, Hugging Face TGI, etc.) during penetration tests and attack surface discovery. Uses HTTP-based service fingerprinting to identify server infrastructure.
• a2a-scanner - Scan A2A agents for potential threats and security issues

AI-Assisted Defensive Security

• Claude Code Security Review - An AI-powered security review GitHub Action using Claude to analyze code changes for security vulnerabilities.
• OpenCodeReview - AI-powered code review CLI from Alibaba with deterministic pipelines + LLM agent, line-level precision, and built-in security rulesets (NPE, thread-safety, XSS, SQL injection).
• deepsec - Agent-powered vulnerability scanner by Vercel Labs for finding hard-to-spot issues in large codebases using coding agents. Supports parallel scanning, PR diff review, and CI/CD integration.
• VulnHunter - Agentic AI security scanner that hunts exploitable vulnerabilities like an adversary, proves them with executable PoCs, and proposes test-first fixes. Maintained fork of Capital One's VulnHunter, rebuilt to run on any agent harness with sandboxed exploit validation.
• defending-code-reference-harness - Reference implementation for autonomous vulnerability discovery and remediation using Claude. Includes threat modeling, scanning, triage, and patching skills.
• Visa Vulnerability Agentic Harness - Agentic SAST pipeline with 11 stages from detection through LLM-powered remediation and adversarial validation. Outputs SARIF, integrates with Claude Code, Copilot, and Gemini.
• GhidraGPT - Integrates GPT models into Ghidra for automated code analysis, variable renaming, vulnerability detection, and explanation generation.
• IDAssist - AI-Powered Reverse Engineering Plugin for IDA Pro. Integrates LLM-powered analysis into IDA's interface with semantic knowledge graphs, RAG document search, and support for multiple LLM providers (OpenAI, Anthropic, Ollama, LiteLLM). Analyzes functions, suggests renames, answers questions about code.
• REA - Local CLI and MCP tools for agent-assisted reverse engineering of native binaries, JavaScript/Electron apps, and .NET assemblies, with evidence provenance and analysis limitations; native deep analysis uses separately installed Hopper, Ghidra, or IDA.
• ThreatForest - Agentic threat modeling platform built on Strands framework. Autonomously generates attack trees from repositories, maps attack steps to MITRE ATT&CK techniques, and produces actionable mitigation recommendations.
• claude-grc-plugin - Claude Code plugin that turns Claude into a senior GRC analyst. 72+ reference files covering 15 frameworks (NIST 800-53, FedRAMP, ISO 27001, SOC 2, etc.), 24 slash commands, and deep domain knowledge for federal and commercial compliance work.
• Vigil SOC - A comprehensive open-source security operations platform for AI agents, enabling real-time monitoring, threat detection, and incident response for AI-powered environments.
• AiSOC - Open-source, self-hostable AI-powered SOC that ingests security events, correlates them, runs autonomous AI-driven investigations via LangGraph, and surfaces results in a unified console. Features full agent decision audit trail, public eval harness in CI, and 52 first-party connectors. MIT licensed.
• Buttercup - Trail of Bits' Cyber Reasoning System (CRS) built for DARPA's AI Cyber Challenge — AI-assisted fuzzing (oss-fuzz based) to discover vulnerabilities in C and Java repositories, then a multi-agent LLM patcher that analyzes and repairs them automatically.

Privacy & Confidential Computing

• Python Differential Privacy Library
• Diffprivlib - The IBM Differential Privacy Library
• TenSEAL - A library for doing homomorphic encryption operations on tensors
• SyMPC - A Secure Multiparty Computation companion library for Syft
• Cloaked AI - Open source property-preserving encryption for vector embeddings
• dstack - Open-source confidential AI framework for secure ML/LLM deployment with hardware-enforced isolation and data privacy
• PrivacyRaven - privacy testing library for deep learning systems
• PLOT4ai - Privacy Library Of Threats 4 Artificial Intelligence — A threat modeling library to help you build responsible AI
• OpenDP - Core library for differential privacy algorithms from the OpenDP project — used to build privacy-preserving ML training pipelines
• Presidio - Context-aware PII detection, redaction, masking, and anonymization framework for text, images, and structured data with customizable NLP and rule-based recognizers

Data & Supply Chain Security

• datasig - Dataset fingerprinting for AIBOM
• OWASP AIBOM - AI Bill of Materials
• Trusera ai-bom - AI Bill of Materials — discover every AI agent, model, and API in your infrastructure
• model-transparency - Sigstore-based tooling for signing and verifying ML model provenance, enabling supply chain integrity for AI artifacts

Agentic AI Security Skills

• Elastic Agent Skills - Collection of skills for Elastic's AI assistant, enabling natural language security investigations across logs, traces, and threat intelligence
• Ghost Security Skills - Agent application security (appsec) skills and tools for Claude Code
• tm_skills - Agent skills to help with Continuous Threat Modeling
• Trail of Bits Skills Marketplace - Trail of Bits Claude Code skills for security research, vulnerability detection, and audit workflows
• Semgrep Skills - Official Semgrep skills for Claude Code and other AI coding assistants. Provides security scanning, code analysis, and vulnerability detection capabilities directly in your AI-assisted development workflow.
• claude-bug-bounty - Claude Code skill for AI-assisted bug bounty hunting. Automates reconnaissance, IDOR, XSS, SSRF, OAuth, GraphQL, and LLM injection testing with 4-gate validation checklist and report generation.
• Anthropic Cybersecurity Skills - 734+ structured cybersecurity skills for AI agents. MITRE ATT&CK mapped, agentskills.io standard. Compatible with Claude Code, Copilot, Codex CLI, Cursor, and Gemini CLI.
• llm-sast-scanner - SAST skill for AI coding agents (Claude Code, Codex, etc.) with structured vulnerability detection across 34 classes. Features source-to-sink taint analysis, Judge verification for false positive reduction, and 99%+ precision/recall on benchmarks.
• sast-skills - Collection of agent skills that turn your AI coder into a SAST scanner
• pentest-ai-agents - 31 Claude Code subagents for offensive security. Specialized AI subagents for recon, web, AD, cloud, mobile, wireless, social engineering, payload crafting, reverse engineering, exploit chaining, detection engineering, forensics, and report generation. Tier 2 agents can execute tools directly with approval gates.
• Mantis Skills - Google's decoupled, sequential, security-focused pipeline of agentic AI skills for autonomously reviewing, deduplicating, validating, reproducing, and patching vulnerabilities across codebases of any scale. Features a multi-stage pipeline (architecture analysis → threat modeling → research → review → reproduction → patching → calibration → reflection), built-in sandboxing, and a continuous learning loop that adapts across iterative runs. Supports RTL hardware, IaC, ML pipelines, and compiled binaries.
• USAP - 80 cybersecurity skills + 13 orchestrator agents with a typed 11-field output contract, an enforced resolvable-evidence gate (no verdict without a resolvable source), connector-agnostic MCP routing to downstream security tools, and human-approval gating for every mutating action. MITRE ATT&CK + NIST CSF 2.0 mapped. Runs in Claude Code, Cursor, Codex, and Gemini.
• reverse-skill - Cybersecurity skill router for AI coding clients (Claude Code, Codex, Cursor, OpenCode). Routes APK/binary/JS reverse engineering, pentest, malware, CTF, and firmware tasks to the right methodology with 43 routing rules, on-demand toolchain bootstrap, and a 173-case routing benchmark.

Security-Focused AI Models

• VulnLLM-R-7B - Specialized reasoning LLM for vulnerability detection. Uses Chain-of-Thought reasoning to analyze data flow, control flow, and security context. Outperforms Claude-3.7-Sonnet and CodeQL on vulnerability detection benchmarks. Only 7B parameters making it efficient and fast.
• Foundation-Sec-8B-Reasoning - Llama-3.1-FoundationAI-SecurityLLM-8B-Reasoning is an open-weight, 8-billion parameter instruction-tuned language model specialized for cybersecurity applications. It extends the Foundation-Sec-8B base model with instruction-following and reasoning capabilities.
• Antares (1B & 350M) - Agentic models for vulnerability localization by fdtn-ai. Available in 2B and 0.4B parameter variants, designed to autonomously identify and localize vulnerabilities in code.
• CyberSecQwen-4B - _4B-parameter CTI specialist fine-tuned from Qwen3-4B-Instruct-2507 for cybersecurity threat intelligence.
• Meta-SecAlign-8B / Meta-SecAlign-70B - Security-aligned Llama models fine-tuned to resist prompt injection attacks, maintaining instruction hierarchy even under adversarial inputs
• Lily-Cybersecurity-7B - Cybersecurity-tuned 7B chat model optimized for security analysis, vulnerability explanation, and threat intelligence tasks.

Cybersecurity-Tuned Text Generation

• ZySec-7B / SecurityLLM - 7B cybersecurity-specialized chat model trained for security Q&A, threat analysis, and compliance guidance.
• SecGPT (Clouditera) - Open-source cybersecurity-tuned LLM (1.5B/7B/14B) fine-tuned on Chinese and English security corpora for threat analysis, CTI extraction, and security Q&A

Safety Classifiers & Prompt Injection Detection

• Llama-Guard-4-12B - Meta's latest multimodal safety classifier for detecting harmful content in LLM inputs and outputs across text and image modalities.
• Llama-Prompt-Guard-2-86M - Lightweight 86M parameter model from Meta for detecting prompt injection and jailbreak attempts in production LLM pipelines.
• ShieldGemma-2B - Google's 2B parameter text safety classifier for detecting harmful content, built on the Gemma architecture.
• Shieldstral-1.0-3B - Mistral's 3B-parameter policy-adaptive multimodal safety classifier. Evaluates text, image, or text+image content against a natural-language safety policy as a single yes/no pass, re-targetable to new policies at inference without retraining, multilingual.
• DeBERTa Prompt Injection Detector v2 - Protect AI's DeBERTa-v3-base fine-tuned for prompt injection detection, widely used in production LLM guardrail pipelines.
• Prompt Injection Sentinel - ModernBERT-large model fine-tuned for prompt injection and jailbreak classification with low false-positive rate.
• Nemotron 3.5 Content Safety - NVIDIA's 4B-parameter multimodal content safety model unifying text+image input guardrails, multilingual support (35+ languages), customizable enterprise policy enforcement, and auditable reasoning in a single inference call. Successor to Nemotron 3 Content Safety.
• BrowseSafe - Specialized security model for detecting prompt injection attacks in AI browser agents. Achieves 90.4% F1 score on BrowseSafe-Bench, optimized for real-time async classification of raw HTML content.
• safe-laya - 421M Apache-2.0 prompt attack classifier (BENIGN/PROMPT_INJECTION/JAILBREAK/HARMFUL_REQUEST) for local, private guardrails. ~30 ms inference, signed checkpoints with CycloneDX AI-BOM, fully reproducible pipeline.

Domain-Adapted Security Language Models

• ATTACK-BERT - Sentence-transformer model for mapping security text to MITRE ATT&CK techniques.
• CySecBERT - BERT model adapted for cybersecurity and CTI tasks through domain-specific pre-training.
• VulBERTa - RoBERTa-based model pre-trained on large C/C++ code corpora and fine-tuned for vulnerability detection. Outperforms prior BERT-based approaches on multiple CVE benchmarks
• Nemotron Content Safety Reasoning 4B - NVIDIA's 4B reasoning-based content safety classifier built on Gemma-3-4B. Enforces custom user-defined safety policies with explainable verdicts, plus a low-latency reasoning-off mode.
• Llama-3.1-Nemotron Safety Guard 8B v3 - NVIDIA's multilingual content safety guard model (Llama-3.1-8B base) classifying prompts and responses as safe/unsafe with violated categories from a user-supplied taxonomy. 9 languages supported, 20+ zero-shot.
• SecRoBERTa - RoBERTa model trained on cybersecurity text for improved performance on security domain tasks.

Datasets

• SafetyPrompts - Curated collection of safety-relevant prompts for evaluating LLM safety and security properties.
• Do-Not-Answer - Dataset of prompts that responsible LLMs should not answer, for safety evaluation and red teaming.
• JailBreakV-28K - Large-scale dataset of 28,000 jailbreak prompts for benchmarking LLM safety.
• CL4R1T4S - Leaked system prompts from ChatGPT, Claude, Gemini, Grok, Perplexity, Cursor, Lovable, Replit, and other major AI tools. Largest known collection of production system prompts for transparency and attack-surface research.
• LEAKHUB - System Prompt Leak Leaderboard — community platform for tracking and ranking system prompt leaks across AI products.
• Leaked System Prompts - Collection of leaked system prompts from commercial AI tools — useful for understanding real-world prompt engineering and attack surfaces.
• SemGuard Arabic Security Dataset - First validated Arabic/Arabizi LLM security dataset — 807 examples across 7 threat categories, validated via 3-judge LLM-as-Judge pipeline (GPT-4o, Grok-4, Llama 3.3 70B), Fleiss' κ=0.839. Includes a 527-example inter-judge disagreement corpus.
• Jailbreak LLMs Dataset - Dataset of 15,140 ChatGPT prompts from Reddit, Discord, and websites for jailbreak research (CCS 2024).
• WitFoo Precinct6 Cybersecurity - Sanitized production SOC dataset (8.4 GB): incident records, IDS/logs from real captures with labeled attack chains, suspicion scores, and provenance graphs for SOC ML and detection research.