AI guardrails & agent security
3,625 agents · most starred first
AI guardrails & agent security, page 10 of 37
Local, offline safety net for AI coding agents — records every file change and reverts anything (one file → a whole rm -rf), and alerts on secret reads. Claude Code · Codex · Cursor. Zero deps · MIT.
Offline-first, precision-first security linter for SKILLS, MCP, plugins, configs and other AI infrastructure.
🐺 Cerberus — local-first security gateway for AI coding agents. Intercept, risk-score & human-approve every tool call (Claude Code, Codex, Cursor, Cline).
Trust and verification for the skills your AI agent installs.
A syscall-level security supervisor for AI agents.
MLSecOps Practical Reference Guide, open-source AI and ML security handbook.
SDL, tools, runtime, and a reference backend for reproducible agentic environments.
Governance-as-code for AI coding agents: author a policy once, enforce it on every agent (Claude, Copilot, Cursor, Codex) via git hooks, CI, and native controls.
Shared canonical names for agent-governance primitives. Maps what each system calls its signals to one reference vocabulary, with per-system crosswalks, so verifiers and tooling compose across ecosystems.
Approve agent sudo command requests
Find text that hides from humans but speaks to machines. Defensive scanner for indirect prompt injection in untrusted content.
Secrets management for AI agents: approve API keys and credentials on your phone with biometrics, delivered end-to-end encrypted into the target process. Plaintext never enters the chat, logs, or relay.
Self-hosted credential broker for AI agents: agents ask for an API call by credential name; a policy engine decides; secrets never reach the agent.
Codex 保安:OpenAI Codex 本机安全审计。实时发现高危命令、凭据访问、提示注入与投毒篡改;只读日志,不代理、不装证书。
Zero-infrastructure secrets manager with MCP server for AI agents. Free and open source.
MCP server for an agent's autopilot: scopes, a daily SOL spend cap, and the propose, execute and undo loop.
Belay is an open-source, local-first security layer for AI coding agents (Claude Code, Codex, Cursor, OpenClaw, Hermes Agent and MCP) that blocks dangerous commands, secret leaks, and prompt injection at the tool-call boundary in under 100ms — no LLM in the decision path by default, no cloud, no phone-home.
Transparent MCP security proxy: tool metadata integrity, call validation, and response analysis. Apache-2.0.
UIBC Core — an executable protocol for bounded-lifecycle autonomy:registry, evidence, gates. Reference implementation + conformance tests. Theory SSOT: zhaoxinghua09-cell/lgd-theory. 代码 Apache-2.0;理论文本保留所有权利。
Secrets that refuse to be stolen. Sealed API keys that decrypt only on your allowlisted repo — and raise a security error if anything tries to print, iterate, or read them. Every access attempt alerts the owner.
Adversarial testing for LLM agents: injection, jailbreaks, tool abuse, measured
Your local AI agents talk to each other. You should know what they're saying. A terminal-first guardian that mediates every MCP call, scores it for risk, and asks before anything dangerous happens.
Ask Key: the Mac key manager for AI agents. API keys stay encrypted on your Mac; Claude Code, Codex and Cursor find them over MCP, you approve each use, and the value goes only to the program, never to the AI. | 请旨:给 AI 用的 Mac 密钥管家。AI 自己找得到密钥,普通密钥直接放行,敏感密钥由你审批,密钥只交给目标程序,AI 看不到明文。
An open specification for runtime authority control in autonomous AI agent systems
Action Control for autonomous agents. Check posture. Gate risky actions. Prove execution. Fits your existing agent stack.
Auditable CSPRNG-backed credential generator for AI agents and machine-readable pipelines. Stable JSON schema, NIST entropy floors, SLSA 3 + cosign signed releases. CLI, Go/Python/Rust libs, MCP server.
Embeddable AI action firewall for MCP servers and APIs
prompt.fail explores prompt injection techniques in large language models (LLMs), providing examples to improve LLM security and robustness.
Trust infrastructure for AI agents — constitution enforcement and cryptographic receipts. Python SDK.
Reverse CAPTCHA for AI agent verification. Prove you're a bot, not a human.
[NeurIPS 2025 D&B] Dynamic Risk Assessment for Offensive Cybersecurity Agents
MoralStack is a governance and safety layer for LLM applications. It analyzes user requests before generation, evaluates risk and intent, and decides whether the AI should answer normally, answer safely, or refuse. The goal is to make AI systems more auditable, controllable, and reliable in sensitive or regulated contexts.
Deterministic tool-call guardrails for pi — enforce rules with before-tool hooks instead of prompts
Cordon provides the guardrails for corporate teams adopting code generation tools.
CI/CD compliance scanner for AI agents. EU AI Act, DORA, ISO 42001. GitHub Action for automated compliance checks on every PR.
Fine-grained permissions for AI coding agents
Ratify Protocol™. Open cryptographic protocol for delegated authority: prove an AI agent is authorized to act on another party's behalf, verifiable offline. Hybrid Ed25519 + ML-DSA-65 (FIPS 204). SDKs in Go, TypeScript, Python, Rust, and C.
AI agent security and governance platform for the full lifecycle. Scan before you ship. Govern and block at runtime. No Azure required.
Open-source API key firewall for AI agents. Proxy gateway with per-token rate limits, budget caps, model restrictions, and instant kill switch. Agents never see the real key.
Security scanner for AI agent SKILL.md files. Static analysis, LLM behavioral prediction, and Docker Sandbox execution.
Python Web Security: OWASP Top 10 with Agentic AI Course
Argorix Lang is a compiled language for secure, verifiable AI-agent communication.
Open governance framework for autonomous AI agents. Classify decisions. Enforce policies. Prove compliance.
Every AI skill, independently scanned. Public, open-source trust scoring for skills, MCP servers, hooks, and plugins across every agent platform. Apache-2.0. saferskills.ai
The open protocol for AI agent identity. DNS-anchored. Post-quantum native. Governed by TrustLayer Foundation A.C.
Policy as code for AI agents: rules are Markdown files in git, enforced before every tool call. A Rust proxy and tool-call gates for 43 harnesses (Claude Code, Cursor, Codex, LangGraph, CrewAI and more) refuse dangerous commands, redact secrets and stop runaway loops. Open source, runs locally.
Sane policies for insane agents
Open specification and reference verifiers for Seal records: ATTESTATION-v1 for AI model calls, ACTION-v1 for agent actions, and GRANT-v1 approvals (draft). Ed25519, canonical JSON, checked offline against a published key.
X-ray for DeepSeek Harness plugins: declared capabilities vs actual behavior. Registry + static scanner + badges.
[SEC 2026] Confundo: Learning to Generate Robust Poison for Practical RAG Systems
AI trust-boundary gateway for OpenAI, Anthropic, and Gemini.
Find what is wrong with the AI agent configuration your repository already has.
Cryptographic receipts that prove an AI output was not altered. Offline verification in 21 lines, no server or account. Ed25519 + SHA-256, 128-byte Rust core, optional ML-DSA-65 post-quantum signatures. Python and Rust.
AI agent safety shield. Intercepts dangerous commands, requires password verification, sends alerts. Works with OpenClaw, Claude Code, Cursor, and any agent platform.
Give agents capabilities. Keep credentials local. A local relay that gives AI agents, scripts, and automation controlled access to HTTP, SSH, and LLM APIs — without exposing real URLs, passwords, or API keys.
A runnable Docker Compose lab demonstrating the three structural failures of observe-and-detect agent governance. No API keys, no cloud accounts, 60 seconds to run.
Trojan Hippo: Weaponizing Agent Memory for Data Exfiltration — benchmark, adaptive red-teaming, and defenses for persistent memory attacks on LLM agents (arXiv:2605.01970)
Compile agent policy prose into deterministic verdict programs: narrow evidence questions for the model, the verdict computed in code. Install: npm i -g jev-compiler
Open Revocable Key Standard: self-identifying API keys with a discoverable, unauthenticated revocation endpoint. Kill any leaked key in seconds — no security@ email threads, no per-vendor integrations.
Jev agent authorization for MCP tool calls: Kinde identity and permissions plus Jev's typed, calibrated decisions, checked server-side before every call runs
Run Claude Code or Codex next to your real secrets: the AI sees placeholders, you see the real values.
AI Agent's Omamori — protect your system from dangerous commands executed via AI CLI tools
Local-first guardrails evaluation server for LLM apps. Scans inputs, outputs, and tool calls against a YAML policy and returns allow / block / redact. No database, runs entirely on your machine.
AI writes. SPARDA proves. A deterministic, offline gate that catches when an AI edit removes a guard, exposes a route, or breaks an invariant — no API key, right in the agent edit loop.
Manage file-based identity credentials with CredAgent (Alibaba Cloud's open-source credential protection software), allowing access only by authorized programs/scripts and preventing AI agents or unauthorized programs from leaking credentials.
Runtime security monitor for AI coding agents
Bast is a free, open-source CLI built to bring security to AI-powered terminal operations. It integrates with Bastio AI Security Gateway to protect your data before it reaches the LLM.
Define what your agent can't do. Because if it gets compromised, those limits are all you've got.
The last line between AI agents and your users. Maze-trace CAPTCHA with behavioral biometrics — invisible to humans, impenetrable to agents.
AI-safe secret manager CLI for Claude Code, Cursor, and other AI agents. Local-first, encrypted, no cloud.
The work visa for AI agents. A high-performance Rust sidecar that deterministicly blocks unauthorized actions before execution.
An initiative to build a fair and sustainable AI ecosystem by identifying and crediting open-access creators whose work shows strong similarity to AI-generated content.
Aira is a series of chatbots developed as an experimentation playground for value alignment.
An AI text classifier 🔎 Which can detect plagiarized text generated from ChatGPT or BARD.
Secure OpenAI interactions with confidence - safeguard your OpenAI API keys and boost your app security
Legacy predecessor to VECTOR. This repo is retained for reference only. Start here instead → github.com/Myth727/VECTOR
MCP Kill-switch for AI agents. Validates infrastructure operations before execution. Fail-closed. Evidence-backed.
The Missing sudo Command for AI Agents.
LEASH — Lightweight Encrypted Agent Secret Handling. A proposed companion standard to MCP for secure, auditable, human-controlled secret access by AI agents.
Agent Policy Engine is an AI agent enforcement runtime (PEP) that prevents untrusted data from becoming executable authority in AI agents
Runtime control of LLM agent behaviors through activation steering vectors. More calibrated than prompting.
Security scanner for AI agent skills and MCP tool bundles — prompt injection, IOC matching, malware detection, ML classifier
Policy-as-code enforcement and observability for MCP tool calls. Wraps AI agent sessions with cryptographic integrity checks, argument-level CEL policies, and a full audit trail.
An LLM-based content moderator. Firefox extension to block webpages unrelated to work, based on page title and URL. Local LLMs with Ollama and Langchain to ensure your browsing history never leaves your device, for complete privacy. Google Gemini also supported.
Agent Identity Protocol -- verifiable, delegable identity for AI agents across MCP and A2A. IETF Internet-Draft. PyPI: agent-identity-protocol
A modular LLM skill for epistemic operations and noetic analysis: a cognitive-security framework (SOC) for classifying discourse and diagnosing orientation, deformation, and concealment; routing engagements through matched case modules and tactics, techniques, and procedures (TTPs), grounded in sound reason and normative disposition.
MCP-style tool-use security playground with permission policies.
Codex 保安: local GUI guardrail and monitoring for Codex
Open-source security firewall for MCP servers. Scan for OWASP MCP Top 10 vulnerabilities, enforce runtime policies on AI agent tool calls, and generate compliance audit logs.
Open-core runtime for Agentic Control Plane: identity, policy, limits, routing, PII redaction, and audit as npm modules.
Scan your LLM-powered applications for authorization gaps, leaked credentials, missing rate limits, prompt injection risks, and other security issues — before they reach production.
The open framework for governed AI development: EU AI Act compliance, PII redaction, MCP proxy, prompt-injection firewall. Python + Rust, Apache 2.0
Open-source control plane for AI agents that take real actions: policy-dictated lifecycle controls, approvals, durable execution, audit receipts, and observability.
The proxy your local agent calls out through. Local freedom, policy-gated egress, one-keystroke approval. — https://trollbridge.dev
Modular multi-model generator of controllable AI news stimuli for misinformation research - agent-based simulation seeding, red-teaming & human-perception studies. GPT-4/LLaMA/Mistral/DeepSeek, multilingual, full provenance. Stimuli are evaluated by human raters in JudgeGPT.
The open standard for cryptographic identity and trust verification of autonomous AI agents.
Superseded by openclaw-acp-plugin. Policy layer for OpenClaw: identity, scope, rate limits, audit.
Corpus poisoning attacks against RAG-based cybersecurity/IoT threat-intelligence assistants, with a lightweight, edge-deployable provenance defense. Evaluated on TF-IDF and dense-embedding retrievers over a CVE/MITRE ATT&CK/vendor-advisory knowledge base.
Runtime security enforcement for autonomous AI agents.