AI guardrails & agent security
3,625 agents · most starred first
AI guardrails & agent security, page 14 of 37
🦞 Your agent's exoskeleton. Immutable Cedar policy boundaries for MCP tool access.
Silent dependency injection through AI documentation pipelines. 240 isolated Docker runs proving Context Hub's zero-sanitization MCP server lets poisoned docs compromise developer projects without warning.
CORD — Constitutional AI safety engine for autonomous agents. Hard blocks, risk scoring, real-time audit. 482 tests.
Antigravity. Claude-code. 🇬🇧 Zero-dependency Node.js CLI to statically audit third-party AI Skills for malicious code patterns before local execution. | 🇪🇸 CLI Node.js sin dependencias para auditar estáticamente Skills de IA buscando código malicioso antes de ejecutarlos.
Human-in-the-loop governance middleware preventing AI hallucination at the Git level.
A free, open-source LLM security gateway. Drop it between your app and any LLM provider to get rate limiting, audit logging, token tracking, and Prometheus metrics with no code changes to your app.
Memory defense for AI agents — stops MINJA, AgentPoison, and MemoryGraft attacks. Zero dependencies.
Transparent proxy service for adversarial prompt injection and SEO
Auth4Agent is a decentralized identity and authentication system designed for autonomous agents, AI systems, services, and machine-to-machine infrastructure.
Operationalizes PM insights through working agents grounded in GRC best practices. Provides prompt libraries and tools to identify governance and compliance risks before scaling programs, analytics initiatives, or AI systems.
A security-first plugin that places a "Human-in-the-Loop" gatekeeper between the AI and high-risk actions.
JS/TS SDKs for Handlebar governance. Integrate into your favourite agent builder
Jupyter notebooks for testing Prisma AIRS AI Runtime with your LLM
A runtime governance, security and execution control layer for Retrieval-Augmented Generation (RAG) systems
Retrieval Pivot Attacks in Hybrid RAG (Vector → Graph)
AI agent governance layer , sign, monitor and control every agent action. EU AI Act · ANSSI · NIST ready.
Governance layer for AI coding agents — every action classified, gated, and cryptographically logged. 751 tests. 3 dependencies. Zero databases. #rai-verify-2026
The Nervous System - LLM Behavioral Enforcement Framework. Anthropic built the brain. This is the nervous system.
A secure email gateway MCP server that protects AI agents from prompt injection attacks hidden in emails.
Adversarial evaluation framework for embodied and agentic AI — failure-first methodology, jailbreak corpus, VLA red-teaming, and policy research.
Open protocol for runtime budget and action authority over AI agents
Documentation for Cycles — AI agent governance, runtime budget, action authority, MCP integration
Prompt-injection defenses for Claude Code. A PreToolUse Bash hook blocks compositional credential-exfiltration shapes (secret read plus network, env dump to network, remote script to shell, reverse shells). A sanitizing MCP server wraps untrusted URLs and files in sentinels, strips invisible unicode, flags jailbreaks.
Demonstrate black-box adversarial attacks on LLM tool-calling to bypass safety and execute remote code using CMA-ES optimization.
AI agent platform with 22 security controls. Signal + WebChat + Discord. Powered by Anthropic Claude.
Governance proxy for AI agents — per-agent budgets, cost tracking, loop detection, and policy enforcement. Agents never hold real API keys.
The guardian layer for AI agents — identity, secrets, audit via MCP. Gate + Vault + Watch.
Local-first AI security gateway for AI Agents. Intercept, audit and govern Claude Code · Codex · Cursor · OpenClaw traffic — 3 monitoring tiers × 8 defense layers, zero cloud.
AI-Native Post-Authentication Runtime Control Plane. LLM-driven Zero Trust security platform for Spring Boot applications. Continuously evaluates authenticated runtime behavior with 5-action dynamic control (ALLOW · CHALLENGE · BLOCK · ESCALATE · PENDING).
🛠️ Build a collaborative framework for pricing strategies using AI, enhancing decision-making through real-time data analysis and human insight.
Append-only event kernel with Ed25519-signed Merkle checkpoints. Every AI action gets a verifiable receipt.
Agentic SpendGuard — audit-chain spend control for LLM agents. KMS-signed decisions, Stripe-style auth/capture ledger, operator approval, multi-tenant. Adapters for Pydantic-AI, LangChain, LangGraph, OpenAI Agents SDK, Microsoft AGT.
Zero-config static security scanner for AI agents. Detects 20 vulnerability categories mapped to MITRE ATLAS and OWASP LLM Top 10.
Experimental Python Trust Kernel for agent interoperability, authorization, budgets, durable governance and verifiable evidence. Bounded protocol support; production qualification pending.
VELLA = Verifiable Evidence Ledger & Logic Authority; deterministic pre-execution authorization and signed proof for AI agents and automated workflows.
Safe-by-default AI agent. Gates every action through Authensor before it executes.
Think like an attacker. Ship like a defender. — Open-source SKILL.md packs for hardening LLM systems: prompt injection, RAG security, tool execution, data leakage & governance. Framework-aware (LangChain, LlamaIndex, Semantic Kernel) and language-portable. Built for engineers who ship to production.
🚪 Governance gate for AI agents. Enforce policies before deployment: request contracts, operational safeguards, identity boundaries, action budgets. CLI tool with YAML config, JSON/text output, CI/CD integration.
Resonance frequency layer with ERE verification
Cryptographic identity for AI agents: Ed25519 keypairs, DPoP request signing, delegation chains.
Governance control plane for AI agents — policy enforcement at the tool-call, spawn, delegate, plan, and budget layer. OPA-backed, Apache-2.0.
Zero Trust agent runtime for auditable AI-assisted engineering and research.
streamlit region:us
docker streamlit region:us
Reversibility-aware safety signals for LLM tool-use agents | estimating action recoverability before irreversible operations: (delete/send/pay/commit) when rollback is unavailable.
Fail-closed authorization gateway for agentic RAG — capability-label pushdown, deny-aware barriers, per-turn agent-context revalidation.
Igris Guardian — a capability-bounded prompt-injection firewall (verdict-only, self-protecting, universally deployable)
Research on why autonomous AI agents don't know when to stop, and three engineered fixes.
MCP & Claude Code security scanner — threat-models plugins, MCP servers, hooks, skills & connectors with an LLM before you trust them. Catches prompt injection, tool poisoning & credential exfiltration.
A deterministic gate for AI-proposed actions: the model suggests, your constraints decide, nothing executes unapproved.
Cryptographic Harness for Uncensorable AI Agent Communication. Post-quantum verification architecture for autonomous AI agents. Fail-closed network policy. NTRU-Encrypt KEM at λ=128. 17/17 leaks fixed. The fence that holds.
Security scanner for AI agent skills and MCP servers. Detects secrets, CVEs, supply chain attacks, and prompt injection in SKILL.md files before they're installed. Pre-install gate for Claude Code, OpenClaw, PicoClaw, NanoBot, FamClaw, and CI/CD pipelines. Single Go binary, MIT licensed.
Runtime security and policy enforcement for tool-using autonomous AI agents
agent-Shield sits silently between your AI IDE (Cursor, Claude Desktop) and your computer. It continuously watches the Model Context Protocol (MCP) tool requests, blocking dangerous commands and API key leaks in real-time.
Open-source MCP-OAuth conformance scanner + IETF reference implementation of draft-ietf-oauth-identity-chaining-14 and draft-ietf-oauth-transaction-tokens-08. CLI + GitHub Action + hosted scanner + public registry. Apache 2.0.
Zero-trust runtime security & governance layer for autonomous AI agents — prompt injection firewall, PII masking, real-time policy enforcement.
Benchmark LLM jailbreak resilience across providers with standardized tests, adversarial mode, rich analytics, and a clean Web UI.
PSA is a multi-classifier behavioral analysis engine for LLM responses. It classifies every sentence into behavioral postures, then derives metrics from the posture sequences to detect adversarial stress, sycophancy, hallucination risk, persuasion techniques, and input pressure — in real time.
The most comprehensive LLM + MCP security guide i.e. OWASP aligned, real CVEs, actionable checklists
🛡️ 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.
AWS-native AI Security Posture Management for AI agents (Observe • Govern • Defend). Discovery, OWASP/NIST/MITRE posture rules, runtime detection & Bedrock Guardrails enforcement, and AIDR integrations (Security Hub, GuardDuty) — open source, in-account, extensible.
Protect autonomous AI agents from wallet-draining prompt injections and ghost paywalls. Deterministic socket-layer pre-payment audit proxy for x402.
MemGuard - eBPF-based real-time memory poisoning detection and alerting for LLM agents (AgentSight overlay)
PACT: Can Enterprise AI Assistants Be Trusted Under Pressure? A benchmark of whether LLM assistants keep following compliance rules in regulated workplaces when a deadline, a manager, or a pushy user makes breaking them convenient. Paper, dataset, leaderboard, and evaluation harness.
gradio region:us
Interactive Secure Edge AI Governance Playground with MCP, deterministic policy gates, two-person approval, attestation, and safe rollback.
Governed AI-assisted software work: practical control, evidence, and change-boundary tools.
Intent-aware runtime governance and enforcement for autonomous AI agents interacting with APIs, tools, and MCP servers.
TRACE-MAS: Unified cryptographic framework securing multi-agent LLM pipelines against drift, spoofing, prompt injection, and temporal attacks
An MCP tool. Connect your assistant to /mcp and every tool call goes through a broker — scoped per caller, under a credential they never hold, revocable, metered and audited.
⚡️ Answer the accountability question about your organization in ninety seconds: seven questions, three files at well-known paths, and a URL you can paste into a vendor questionnaire. Not released yet.
Monitor your local AI agents. See what files they can reach, what networks they communicate with, and stop them if required. Detect a rogue agent.
A local-first runtime that runs AI agents behind a capability arbiter. Every tool call is gated against a declared manifest, scored on earned trust, and written to an immutable audit ledger.
Governed execution for AI agents, workflows and software.
Post-response grounding check + cross-family adversarial review for LLM chat agents
Vexa Agent Control helps engineering teams govern the AI-agent traffic they route through it, with centrally managed policies and verifiable, per-surface enforcement evidence.
Executed-evidence verification gate for AI coding agents
Check what the user may do before your AI agent does it. Live permission checks against Jira, GitHub, Kubernetes, AWS and 17 more systems.
Catch legal and compliance obligations (opt-outs, privacy requests, legal threats, chargebacks) in customer messages before your chatbot replies. YAML policy packs, Python and TypeScript, HTTP sidecar.
Local secret store for AI coding agents. Agents run commands with your secrets but never see them: a desktop window or your Android phone asks you to approve with a YubiKey touch, Touch ID, or passphrase, grants expire on a timer, and output is redacted. age-encrypted, no server. Linux, macOS, Windows, Android.
A simple web based tool that analyzes password strength in real time and gives instant feedback to help users create stronger more secure passwords
Catch unsafe tool calls before you ship. Test prompt injection in tool-using agents with reproducible traces and benign controls.
Meet Jack: The First Agent for Real Privacy and Security. By separating reasoning from system authority, Jack enforces strict local safety gates, sandboxed math solvers, and real-time streaming IRQs to turn untrusted LLMs into highly secure operational tools.
Verification infrastructure for AI reasoning. Privacy by design, safety through cryptography.
APEX-Harm: Evaluating Agent Safety in Shared Professional Workspaces
Open-source AI security gateway. PII, jailbreak, prompt injection, secret detection - self-hosted.
MSc dissertation (GISMA, 2026): a controlled benchmark of indirect prompt injection, PoisonedRAG and Crescendo against four open-weight LLMs in a RAG pipeline, with three low-cost defenses, exact statistics and an EU AI Act Article 15 mapping.
An invisible, zero-network security proxy for the Model Context Protocol (MCP).
Execution-layer governance for agentic AI: within a modeled action space and registered invariants, unsafe actions are made structurally unreachable — not merely refused. An open architectural hypothesis under public review.
A pre-decoding detection method for system prompt leakage and RAG leakage attacks.
Control autonomous AI agents by enforcing behavior rules to prevent unauthorized actions, improve focus, and boost execution efficiency.
Free Claude Code skill & plugin that makes AI coding agents safe: asks before risky actions, catches hallucinated code, prompt injection, SQL injection & leaked secrets, reviews every change, and writes PDF reports. By Sami Hussain (x8code.com).
A practical Tasfia AI persona prompt for authorized cybersecurity learning, defensive research, CTFs, and secure development.
Assurance methodology for autonomous AI agents in regulated financial services
AgentPin agent pinning protocol, part of the Symbiont Agent Trust Stack
Cryptographic agent identity with Ed25519 anchors, signed action receipts, and scoped trust delegation. One .aid file. Any agent. Any LLM.
The user sees text that doesn't exist = Chrome's font settings + a malicious OpenType font + an AI agent's legitimate permissions. Here's the full chain and why it's hard to operationalize.
MCP governance proxy for AI agents — enforce rate limits, access control, human approval, and audit logging at the protocol level.