AI guardrails & agent security
3,625 agents · most starred first
AI guardrails & agent security, page 33 of 37
Static analyzer for confused-deputy vulnerabilities in MCP, LangGraph, OpenAI Agents, and CrewAI systems
Deterministic validation of AI-agent completion claims against evidence and authority requirements.
Catalog-backed API key loader for coding agents
Signed, verifiable AI-governance measurement card (Ed25519 receipt-spec). Measurement, not certification.
Allowlist-based task packets for Hermes, OpenClaw, and other AI agents
Offline, bounded claim-to-evidence receipts for agent handoffs
EU AI Act Article 12 audit trails for AI agents
Read-only reliability and safety auditor for AI agent repositories
Deterministic schema validation gates and circuit breakers for LLM function calling
Local identity for an agent: wallet, secrets, and optional email
Check what an LLM is given for prompt injection: material at ingest, and the turn it answers
DeepKeep AI Firewall guardrails and hooks for CrewAI
LangChain integration for Keycard: delegated, per-tool-call access with brokered credentials
LangChain middleware for NeuralTrust TrustGuard
A pytest-native security testing framework for LLM applications (OWASP LLM Top 10).
Security testing and authorization-aware retrieval for production RAG systems.
LiteLLM ↔ TrigGuard — PARTIAL callback gate + FULLY_ENFORCED protected_completion path
Instant file protection for AI agent operations - hardlink-based zero-copy backup system
Open detection standard -- like Sigma, but for AI agents. Executable rules for prompt injection, tool poisoning, context exfiltration, and MCP attacks. Shipped in Cisco AI Defense. MIT-licensed.
Fail-closed AI-agent authority verification, delegated capability enforcement, and signed decision receipts
A minimal-privilege proxy for MCP tool calls
A circuit breaker and runtime protection SDK for AI agents.
Runtime governance for AI agents. Conduct Guard as a LiteLLM guardrail — enforce runtime policy on every LLM call routed through LiteLLM.
Pre-execution Decionis policy gates for current AutoGen tools and legacy UserProxyAgent.
Gate CrewAI tool calls on a signed Decionis Decision Dossier before execution.
SOVP to AgenTrust TRACE bridge
Open-source security scanner for AI-powered and agentic applications.
DeepKeep AI Firewall custom capability for Pydantic AI
Watchlight governance hooks for Claude Agent SDK agents: authorize every tool call and sub-agent spawn before it runs, fail-closed. Model turns are not authorized. Independent plugin, not affiliated with Anthropic.
Governance primitives for LangGraph agents with Watchlight: authorize gated actions, validate plans, spawn attenuated sub-agents.
Watchlight governance for Pydantic AI agents: authorize the tools you wrap with guarded_tool() and open every Agent.run as a governed run (nested runs as attenuated sub-agents), fail-closed. Undecorated tools and model turns are not authorized.
AI agent runtime security: context-preserving tokenization, risk scoring, and policy enforcement
CAITLYN Security Guardian Agent — interactive defense agent for LLM agent protection
Defensive LLM agent honeypot & token tarpit — chat widget, PHP2Shell HTTP decoy, Cloudflare Worker, and Node adapters.
Gate every LangChain agent action through Qerexa runtime trust enforcement.
Lightweight SDK for routing AI agent LLM traffic through a guardrails proxy
HeartFlow (心虫) discriminator tool for CrewAI — gate agent output before it reaches users.
Intertrace runtime protection SDK — authorize+permit tool gates and agent-builder adapters
Trust Agent Protocol (TAP) SDK — agent registration, policy evaluation, human review, TAP App passports, audit, and evidence
BTP v7.0.0 Agentic Runtime Protection (ARP) Gateway & MCP Safety Shield - Claude Code, Google GenAI, LlamaIndex, Microsoft AutoGen & Semantic Kernel Middleware, In-Process Tool Invariants, Sub-20µs RAM Gating, and Cryptographic Merkle Vault
Agentic Control Plane governance for Pydantic AI agents. One ACPHooks() capability; ACP decides allow/deny/redact on every tool call.
Traceable text watermarking for AI-agent long-form deliverables — anti-leak attribution
Python support library for the Agent Approval Interchange Specification
Static, pre-deployment analysis of what an agentic system can reach in an AWS account from untrusted model input.
Definitions for agent security tools: public rule sets and risk corpora, scored against outside benchmarks, watched for changes, and preserved after the source disappears.
Official Python SDK for the Agent Enrollment Protocol
Outbound change control for AI agents — gate shell, file-write and HTTP side effects before they happen
Ultra-low latency (<3ms) deterministic security, prompt injection, Polygon OMS x Solana Pre-Sign Guard, ElizaOS Brain CFR, and Universal Modular Escrow on Polygon, Base, Arbitrum, and Solana Mainnet.
Policy-enforced self-edit gateway for coding agents
Human-in-the-loop approval firewall for Python agent tool calls
A firewall for AI agents: block prompt injection, guard tool calls, redact PII/secrets.
Local verifier for agentic data handoffs: envelope, provenance, trust, authority, identity, capability, and audit checks.
Operator-approved digest-only Harness extension for Transfer Verifier.
Embedded AI agent governance peer. Standalone or connected to Agentomy infrastructure.
Mint and sign agent Ed25519 keys (did:key). CLI only; no MCP.
Optional advisory-only AI Agent Handoff extension for Agentic Security Harness.
Check what an LLM is given for prompt injection: material at ingest, tool output, and the turn it answers
Runtime firewall for AI coding agents — gate every tool call, MCP call and skill against a least-privilege policy.
The agentic behaviour layer for Keystone -- declare rules, behaviours and triggers, deploy them in an Engine, and gate any keystone-connect write; every firing is a content-addressed behaviour_ref (Apache-2.0)
Cryptographic runtime receipts for Microsoft AutoGen agents — signed Ed25519 audit evidence for every tool call.
Cryptographic runtime receipts for LangChain/LangGraph agents — signed Ed25519 audit evidence for every tool call.
Cryptographic runtime receipts for LlamaIndex agents — signed Ed25519 audit evidence for every tool call.
Cryptographic runtime receipts for OpenAI Agents SDK — signed Ed25519 audit evidence for every tool call.
Cryptographic runtime receipts for Pydantic AI agents — signed Ed25519 audit evidence for every tool call. CCS (Correctover Conformance Shape) L1 + behavior receipts.
Honest human-in-the-loop consent for AI coding agents: you sign the real argv, not the agent's story. Skill census, MCP sampling deny, inbound fence, and a hash-chained receipt of every approval.
An interface for SSH agent socket providers, which expose an abstract communication channel for an SSH agent to derivepassphrase.
A non-hardened implementation of the SSH agent protocol, for use in testing.
A policy enforcement layer between AI agents and the infrastructure they act on.
Simple & fast trufflehog secrets scanning for LiteLLM
Audit the attack surface of a machine running local LLMs (Ollama, LM Studio, vLLM, llama.cpp...). Stdlib only.
Data-only defensive playbooks and canonical offline advisory policy pack.
Onyx Gate for CrewAI — a verifiable authorization gate for crew tool calls: policy-decided allow/deny before a tool executes, with an audit trail you can re-check.
A tamper-evident, hash-chained audit ledger for AI agent actions.
Autonomous Security & Zero-Trust Governance Guardian for Enterprise AI Agent Fleets by Prema AI Labs
Gate every LangChain agent action through Qerexa runtime trust enforcement.
Evaluation harness for indirect prompt injection and retrieval poisoning resistance in RAG pipelines
LangChain integration for SafePrompt: prompt-injection detection as a callback handler. Validates every prompt flowing through a chain before it reaches the LLM.
Yanez agent authorization: request human approval, verify signed receipts.
The AI Agent Firewall. Enforce signed authority before consequential actions and protect sensitive data at the execution boundary. Free API key required.
The standard for agentic identity on Nostr, plus a thin, dependency-light conformance primitive: the identity types, the act-outward gate predicate (owned + self + verified + funded), and a pure structural conformance checker. ESM + CommonJS.
APort Repository Guard and pre-action authorization guardrails for GitHub, Claude Code, Cursor, Codex, Gemini CLI, Goose, OpenClaw, LangChain, CrewAI, DeerFlow, and n8n
MCP policy gateway and red-team benchmark harness
Authorization and evidence SDK for AI agents: check permissions before every action, record what agents actually did.
Pre-execution gate for CrewAI: check an MCP server or package with RelayShield before an agent connects to it
Exact-action authorization receipt checks for Hugging Face smolagents tools.
Deterministic pre-deployment checks for AI agent systems. Blocks unsafe designs before they ship.
Keep API keys on your phone and lend them to an AI agent's machine — in memory only, never on disk, and only while you allow it.
Detect if code is running in an AI agent or automated development environment
Local deterministic pre-attachment security scanner for MCP and AI-agent tool manifests.
Tiny provider-neutral authorization checkpoint for AI-agent actions: allow / deny / require_approval.
Pre-commit secret scanner. Blocks API keys, tokens, .env files, and private keys from leaking into git. Ships as a skill for Claude Code, OpenAI Codex CLI, Google Antigravity, and Moonshot Kimi CLI, plus a standalone CLI / git hook.
MCP server + live dashboard for AI code governance — OWASP LLM Top 10 (10/10), real-time MCP App UI, 25+ security patterns, Bayesian learning Brain, hallucinated import detection, multi-agent governance. Works with Claude, Cursor, VS Code, ChatGPT, Goose,
Proof of what your AI agents did. Give your agents real power, without losing control of them. The Python client of the agent-custody sidecar.
AIBOM generator for AI agent harnesses (Hermes, OpenClaw) and their model/skill/MCP/hook surface
Operational safety controls for autonomous AI agents - kill switch, heartbeat, circuit breaker
Verifiable, attenuable delegation for AI agents - org-rooted and verified offline
Runtime AgentCreds agent-identity enforcement for tool calls - MCP middleware and agent-to-agent (A2A) verification
Flags AI agent code that persists untrusted tool/external/user content into long-term memory with no sanitization or provenance marker (agent memory poisoning).
A robust, enterprise-grade, drop-in security & safety guardrail layer for AI agents and LLMs.
Runtime control for AI agents - replay any run, prove what happened, veto what must not happen.