AI guardrails & agent security
3,625 agents · most starred first
AI guardrails & agent security, page 7 of 37
A measured tool-call firewall for AI agents: allow, ask or deny before the agent acts. Claude Code mod and hook, MCP proxy. Built on TypeSafe Jev, measured on real sessions with a blind second labeler.
A calm reliability layer for AI agents: thoughtful execution, clear safety boundaries, reusable memory, and one-click installation.
Runtime governance for AI agents. Contracts enforce what tools can do — before they execute.
A Deception Security Layer for MCP Servers. It injects "ghost tools" (fake security-sensitive tools) that act as honeypots.
Code for the API, workload execution, and agents underlying the LLMail-Inject Adpative Prompt Injection Challenge
Static analysis and attack-surface scanner for AI agents — SAST for prompts, tools, MCP servers, and agent workflows. Detects capabilities, prompt risks, tool permissions, and governance gaps before deployment.
Framework-neutral control contract for AI agents: eight interception points, three verdicts, and a conformance kit.
The execution safety layer for AI agents.
Adversarial security testing for AI agents — prompt injection, tool misuse, memory poisoning, and MCP scanning, with CI-ready regression testing and reports.
Static security auditor for AI-agent configuration: Claude Code, MCP servers, Cursor, Codex, Copilot, Gemini CLI and more. SARIF output, OWASP LLM Top 10 mapping.
An ML-powered framework for detecting and counteracting sycophantic spiraling in AI chatbots.
Secure-by-default OpenClaw on Ubuntu, with a verifiable security report
A curated collection of research and techniques on securing agent communication in Large Language Model (LLM)–based agent systems, including protocols, security risks, defense countermeasures, and more.
Lightweight AI safety middleware that protects humans by intercepting self-harm and criminal intent in LLM prompts. Features a 3-stage safety pipeline, MCP server for agents, and automated care responses.
The system of record for AI. Models commoditize, trust doesn’t.
Zero-dependency TypeScript SDK for AI agent governance: policy enforcement, injection detection, tamper-evident audit, and standards mapping (EU AI Act, OWASP, NIST, ISO 42001)
Portable security rules for the action boundary of AI agents
ZAK — Zeron Agentic Kit. Open-source ADK for building, deploying, and governing autonomous cybersecurity agents.
Real-time AI safety guardrails for LLM apps. 10 scanners: prompt injection, PII, harmful content, code vulnerabilities, obfuscation detection. Sub-ms latency. Python + TypeScript SDKs. MCP proxy. Claude Code hooks.
Find, prove and control the AI agents on your machines: discovers every agent and what it can reach, proves policies with Z3 + TLA+, enforces them deterministically outside the model.
ClawGuard is a comprehensive security toolkit designed to mitigate risks associated with autonomous agents, such as OpenClaw and other LLM Agents.
Runtime security platform for LLM agent tool calls on Kubernetes
[ICLR 2026] Official implementation of "ChatInject: Abusing Chat Templates for Prompt Injection in LLM Agents"
Improving AI Systems with Self-Defense Mechanisms
AI-free static security scanner for Claude Code artifacts (Skills, Hooks, MCP configs). Detects data exfiltration, prompt injection, and supply chain risks with deterministic, reproducible results.
A verified runtime for trustworthy AI agents: every claim passes an 8-layer verification pipeline before it counts. Deterministic checks, explainable verdicts, no LLM judging an LLM. Pure C, zero dependencies, MCP-native.
🔪 Open-source safety firewall for AI agents. Intercepts tool calls before they execute, enforces YAML policies, and kills dangerous operations in real-time. Works with OpenAI, Anthropic, LangChain, and MCP. She doesn't guard. She kills.
Declarative guardrails and safety controls for .NET AI agents
Just-in-time API keys for AI agents - and any other process you route through it: the caller only ever sees a placeholder.
The identity provider for AI agents — registry, scoped delegation, in-path MCP enforcement, instant revocation, tamper-evident audit.
Transaction Tokens (RFC 8693) for Kubernetes — seal identity, context, and authorization across multi-hop agent workflows via AgentGateway + ext_authz.
One command to keep secrets out of AI (LLMs). Works with Claude Code, Cursor, Copilot, Windsurf, and any AI coding tool.
[EMNLP 2025 Oral] IPIGuard: A Novel Tool Dependency Graph-Based Defense Against Indirect Prompt Injection in LLM Agents
Cerbere is a self-hosted runtime security layer for AI agents. Every tool call is checked before it executes. No customer data leaves your infra. Open source. Works with MCP, LangChain, CrewAI, Claude Code, Cursor.
A credential-helper framework and agent for Bazel and similar tools implementing the credential-helper spec. authentication, aws, bazel, credentials, gcs, google-cloud, s3
The governance layer for AI agents — monitor reasoning, audit tool calls, and secure the loop through OHTTP privacy routing.
Cryptographically signed policies that constrain AI agent behavior. Same attestation primitives as cilock, applied to agent execution: identity binding via MCP, JWT + signing-key separation, sublayouts for sub-agents.
200 AI agent skills, hardened with targeted behavioral guardrails. Free drop-in replacements.
OpenAI secrets engine for HashiCorp Vault
ThumbGate Pre-Action Checks self-improve from ranked lessons and repeated failures, hard-block detected secret leaks, and block matches in strict mode.
The open-source security layer for AI agents. Deterministic guardrails, PII redaction, and EU AI Act compliance in one line of code.
Open-source EU AI Act compliance scanner. 51 checks across Articles 9-12, 14-15. Drop-in trust layers for LangChain, CrewAI, AutoGen, OpenAI. Local-first, no data leaves your machine.
🔐 The password manager for terminal users and AI agents. Age-encrypted, keyring-cached, MCP-ready. Zero telemetry.
The open-source safety layer for AI agents — block unsafe tool calls, require approval, enforce budgets, audit, replay.
Intent-bound action authorization for AI agents — deterministic, per-action approval a prompt injection can't re-point, with a signed, tamper-evident audit trail.
Authorization infrastructure for AI agents - policy enforcement, signed WASM bundles, per-request user scope, and audit trail for OpenAI Agents, LangChain, Google ADK, and Pydantic AI.
Find AI/LLM security vulnerabilities in your code before attackers do — covers prompt injection, MCP tool poisoning, RAG data poisoning, and more
A local-first permission firewall and approval layer for AI agent tool calls.
Action guardrails for AI coding agents — native Claude Code enforcement, MCP integrations, and Git/CI checks.
The missing governance layer for AI agents. Declare what your agent can do, what requires human approval, and what it can never do - before it runs. Runtime enforced via scyvera, not just documented.
Protect .env secrets from AI coding assistants. Your phone is the key — iPhone Face ID or Google passkey. We also support Windows Hello.
Verify the context you feed your AI agents, against the web or your own documents.
Make destructive AI-agent actions reversible by default — quarantine + audit + human escalation for rm/git destructive operations. Reliability infrastructure, not a sandbox.
⚖️ Kadı — a calibrated judge in front of your frontier LLM. Routine decisions answered locally with a conformal-prediction guarantee (≤ α disagreement with the oracle); everything else escalated and learned from. Zero-dependency Python.
Policy-aware AI permission reviewer for OpenCode V1 and V2. Evaluates tool and command requests, allowing, denying with feedback, or escalating to the user.
Open tools that check claims about AI agents against evidence you pin: an offline verifier with replayable decisions
DeepSeek Harness 破甲深度分支:首尾夹击 + 生成层熔断 + 宿主源码级补丁。基于 YuJunZhiXue/dsh-purge (MIT) 改造,适配 dsh 桌面版 0.2.1-rc.3。仅供学习研究与红队评测。
Plug-and-play terminal security layer for LLM agents. Drop-in gatekeeper that prevents dangerous shell commands. Works with OpenAI, Claude, Gemini & more.
A lightweight Chromium extension that transparently strips out hidden watermark markers from ChatGPT outputs before they ever reach your clipboard.
Whitebox & Blackbox red-teaming framework for LLMs & Agentic AI apps. It analyzes your app's source code to discover tools, roles, and guardrails, then generates new attacks chains across several categories and adapts over multiple multi turn rounds to find vulnerabilities
A police department for your Claude Code agents — a logging-only hook + CLI that audits the main agent and every subagent and reports rule offenses (permission bypass, out-of-scope & credential access, self-permissioning, disallowed tools, redundant, off-task) with quoted evidence. Catch-and-report, never blocks.
Open standard for Provenance & Intent Contracts (PIC) in AI agents. Verify intent, provenance, and evidence before high-impact tool calls.
⚸ Lilith Zero - Security Middleware for MCP tool calls written in Rust.
Demo agents showcasing CapiscIO Agent Guard and MCP Guard — trust badges, identity verification, and tool-level authorization for A2A and MCP protocols
A gui that allows you to bypass content filtering in ChatGPT through the OpenAI autocompletion API for DaVinci-003
Client-side retrieval firewall for RAG systems — blocks prompt injection and secret leaks, re-ranks stale or untrusted content, and keeps all data inside your environment.
The immune system for AI coding agents
Guardrails service for AI agents. Default-deny tool call evaluation with LLM safety analysis, priority-ordered decision matrix, and human-in-the-loop escalations. Session recording, behavioral analysis, MCP proxy, secret redaction, and real-time audit.
The industry-standard Agentic Identity & Inventory Scanner. Automatically inventory autonomous agents (LangChain, AutoGen, CrewAI, PydanticAI) using static analysis, network heuristics, and eBPF. Foundational tool for AIBOM compliance and AgentOps governance.
Agent Receipts — cryptographically signed audit trails for AI agent actions. Protocol spec, SDKs (Go, TypeScript, Python), and MCP proxy.
Auth for AI agents and humans. First-class agent identity, MCP OAuth 2.1, delegation, audit. TypeScript, edge-native, MIT.
Guardrails for AI coding agents — block catastrophic tool calls before they run. Semantic, agent-neutral, near-zero false positives.
Command-line tool for TypeSafe's Jev AI model
Reversible side effects for AI coding agents — snapshot files, hold PostgreSQL writes, and review outbound actions before they become permanent.
Scan a codebase for AI SDKs, MCP servers and agent tools; map risks to NIST AI RMF, ISO 42001, EU AI Act and the OWASP LLM and Agentic Top 10
Claude and ChatGPT now log into your app as your customers. OneHuman decides what they can see and do. Join our community: https://discord.gg/9QX9jTwBq
Request verifiable identities from users or agents, backed by Proof Digital IDs.
A CLI secrets store & dispatch tool built for AI agents. Authy stores encrypted secrets locally and dispatches them to agents with policy-based scoping, short-lived session tokens, and audit logging. No server required.
MCP is being adopted rapidly. Security guidance is lagging behind. This checklist gives security engineers, platform teams, and technical leaders a clear, actionable baseline for securing MCP deployments , whether you're shipping an internal tool or a customer-facing AI agent.
Security Dashboard for OpenClaw AI Agents - intercept, monitor, and control what OpenClaw does on your system.
Pre-execution intent verification for AI agents. Audits what your AI is about to do, not what it says. Zero dependencies, deterministic, hash-sealed.
Allows to red-team your AI systems through adversarial probing. It is simple, effective, and requires minimal setup.
Protect .env secrets from AI coding agents. CLI (Rust) + VS Code extension. AES-256-GCM vault, OS keychain, Touch ID, smart sandbox values. Agents see fakes — you see the real ones.
System-level security for LLM agents: fine-grained policy enforcement on tool calls to defend against indirect prompt injection
Trust metadata for AI agents — a stamp costs ~15 tokens, re-verifying costs 15,000. Agents stamp what they verify; the next agent runs 'akf check' and builds on it. pip install akf
Real-time guardrail runtime for AI agents.
Screens an AI agent's tool calls before they run and tool results before the agent reads them. An MCP proxy plus a hooks adapter for a client's built-in tools.
The firewall for AI agents. Every tool call, every API call: gated by policy, logged, and routed for human approval.
Enterprise Zero-Trust AI Security & Governance Gateway: Real-Time Prompt Injection Defense, Action Scoping, HITL Approvals, Post-Quantum Merkle Audit & EU AI Act Compliance for Autonomous Agent Swarms.
Open-source Prompt Firewall — deflect up to 95% of redundant LLM traffic before it leaves your infrastructure. Documents: https://isartor-ai.github.io/Isartor/index.html
An opensource DevSecOps Layer for your AI agent. Governance, Custom Guard Rails and Observablity at one platform.
Self-evolving minor-user identification agent for anthropomorphic AI interaction, with trigger evaluation, evidence chains, and deployable protection workflows.
Deterministic AI defense framework: immutable input filtering, n-model cryptographic hash consensus, and self-learning adaptive rules. Zero dependencies. Hardware-sealed.
Research MCP, OAuth, security
A transparent proxy system for inspecting, debugging, and modifying any traffic from the agent container, focused on LLM-based coding agents. Works with Claude Code, Codex CLI, Cursor, Aider, and any other tool that calls LLM APIs.
AegisSovereignAI: The Cross-Ecosystem Trust Layer for the Distributed Enterprise. Verifiable Identity, Hardware-Rooted Integrity, and Sovereign AI Governance - from Silicon to Prompt. Unifying AI, Cloud-Native, and Decentralized architectures.
🛡️ Offline Rust security scanner for AI agent tools & MCP servers — MCP, Hermes Agent, CrewAI, LangChain/LangGraph, GPT Actions, Cursor Rules. Catches command injection, credential exfiltration, SSRF & unsafe file access; a typical MCP server scans in <50 ms. CLI, GitHub Action, VS Code, SARIF.
a secret store for ai agents with one rule: the agent never sees the plaintext. get a secret once into the os keychain, then inject it into commands forever. no get, cross-platform, mit
Python SDK for accurate and verifiable agent tool use. Agents verify that answers came from the right source and were not changed. Downstream agents detect 100% of errors and retry to achieve a 50% jump in answer accuracy.
Security Agent Benchmarking and Evaluation Research