ai-agents-security

LLM | Agentic | Security | Operations in one github repo with good links and pictures.

MCTS (Model Context Threat Scanner) is a local-first security scanner for MCP servers -- static and live tool discovery, multiple analyzers, auditable risk scores, and JSON, SARIF, and HTML output. For authors and platform teams; CI-ready, no cloud API.

Open Agent Trust Infrastructure (OATI) — an open standard for verifiable AI agent identity, delegated authority, policy enforcement, and signed action receipts.

Adversarial security review for AI skills, repos, MCP servers, and packages before you install them. A read-only scanner plus a five-persona reasoning pass, ending in a GO / GO WITH MITIGATIONS / NO-GO verdict.

Semantic permissions for AI coding agents. Intenter remembers what commands actually do, not just what they’re called.

🛡️ Official monorepo for SupraWall security plugins and framework integrations. Deterministic zero-trust guardrails for AI agents, including Prompt Injection protection, PII scrubbing, and policy enforcement across Dify, MCP, LangChain, CrewAI, AutoGen, and Vercel AI.

Local-first AI project memory and control runtime for evidence, privacy, model routing, human approvals and cross-harness handoffs.

AegisMCP Enterprise is a dual-layer security gateway and cost firewall designed to sit between AI models (like Claude and Cursor) and execution engines.

LLM Gateway built with LiteLLM for unified access to multiple AI providers through a single OpenAI-compatible API. Features intelligent model routing, failover handling, load balancing, and centralized LLM management.

Policy-as-code admission controller for AI agent skills and MCP tools. SkillBOM, lockfiles, and supply-chain baselines.

Local-first security and cognitive-risk scanner for MCP tools and agent skills. ToolBOM, attack harness, and policy checks.

A framework for building and running your own living GenAI/LLM threat model. Ships with a curated reference catalog, an LLM-driven assessor, and an MCP server

A CLI-only Docker workspace for AI coding agents — Claude Code, Copilot, Codex, Gemini, Kiro — running as a non-root user behind a deny-by-default outbound firewall.

Deterministic policy engine for AI agent x402 payments [Python SDK]

Deterministic policy engine for AI agent x402 payments [TypeScript SDK]

This repository is a local literature collection on agentic AI systems that interact with the web: LLM and multimodal agents for autonomous browsing, browser automation, website traversal, information seeking, web scraping, scraper generation, and web-agent safety/security.

Daedalab is a control layer for AI agents that intercepts tool calls, blocks dangerous actions, enforces spending limits, and keeps a complete audit trail, giving users control over what their agents can do before it happens.

Research prototype exploring how enterprise AI platform teams identify impacted AI agents, run evaluations, and safely redeploy after enterprise changes.