AI guardrails & agent security
3,625 agents · most starred first
AI guardrails & agent security, page 15 of 37
Your AI agent has no idea it's been hijacked. AgentGuard does
Design notes on execution boundaries and responsibility structures for AI systems interacting with the physical world.
A novel approach to mitigating social bias in Large Language Models through a multi-judge pipeline that assesses and filters reasoning steps of DeepSeek-R1.
autonomous agent framework built to detect and mitigate training data poisoning in LLM development. Inspired by the scientific method, the system uses a modular suite of AI agents—hypothesis generators, script builders, critics, and testers—to audit data for adversarial patterns. to improve llm data input. utilizing a crewai framework and eletron
Deterministic execution boundary for AI agents. IFC enforcement at the sink. 5 frameworks. 50 attack vectors. Apache 2.0.
🛡️ The CrowdStrike for AI Agents — Runtime protection, threat detection & security monitoring for LLM agents. Supports LangChain, CrewAI, AutoGen, OpenAI. Python • JS • Rust • Go • Ruby • .NET
ProvnAI builds security infrastructure for autonomous AI agents: runtime enforcement, governed execution, and audit-ready evidence at the boundary between agents and tools.
Scoped, time-bound, cryptographically-signed capabilities for AI agents. Zero dependencies. Sovereign by design.
Secure, zero-knowledge secret manager for autonomous AI agents with human-in-the-loop approval.
An Identity, Security, and Governance Framework for autonomous agents
4th.GRC — A policy-driven Agentic AI governance and orchestration platform integrating PolicyEngine, Skill Registry, RAG Safety Envelopes, Constraint Evaluators, and evidence-based compliance automation.
A self-healing AI orchestration layer that enforces deterministic validation and security over LLM pipelines to prevent hallucinations and unsafe outputs.
AI Agent Security: Otonom yapay zeka ajanlarının ve MCP protokolünün güvenlik riskleri. Excessive Agency, tool poisoning ve Meta'nın "İki Kural" (Rule of Two) savunma çerçevesi.
A collection of resources documenting my research and learning journey in AI System Security.
AI agent credential broker — temporary, task-scoped tokens with automatic revocation. Free for internal use
Deterministic firewall for MCP agent tool calls. YAML policies, hash-chained audit, blast-radius scoring, multi-cloud cost intelligence.
A real-time, zero-trust governance layer that sits between AI agents and the outside world.
Seamless auth flow for the agentic web
System of record for AI tool risk: inventory, policy enforcement, approvals, and audit-ready evidence.
Proxilion is the security layer for the agentic workforce. It turns managed AI agents into governed users by enforcing strict cryptographic boundaries on every API call to SaaS like Google Workspace, Salesforce, or Atlassian.
Benchmark for evaluating safety of AI agents in irreversible financial decisions (crypto payment settlement, consensus conflicts, replay attacks, finality races).
Causal Processing Unit: permission-first engine for cause→commit→execute pipelines (Gate/Incubator/vCML).
Research and Course material based on AI Security Research
Neovim integration for CodeGate
Context Integrity Verification System - система защиты ИИ-агентов от атак на память и контекст, таких как memory injection, prompt injection, tampering и replay.
Claude Code security snippets for safer permissions, sandboxing, and shell guardrails.
Self-evolving recursive AI sentinels for real-time IP protection and prompt injection defence. AI security infrastructure - A token incentivized network.
The full-stack safety layer for AI agents. Budget limits, prompt injection shields, PII filtering, output firewalls, and hooks — in 2 lines of code.
Signed, offline-verifiable evidence for tool-using AI systems.
BETO formalizes the ignorance of an AI. Materializes raw semantic intent into fully traceable, auditable software.
Stop AI agents before they do damage — behavioral sequence detection blocks credential exfiltration, config harvesting, and secret leakage before the network call fires
Semantic mistake-memory and guardrail for AI agents. Stops agents from repeating the same failures using causal graphs and semantic matching. Supports MCP, LangGraph, and in-process Python.
Runtime guardrails for AI agents that enforce token budgets, loop limits, and tool rate limits locally.
Local-first governance framework for AI agents. Tamper-evident audit logs, pre-execution gatekeeping, kill switches. No tokens.
A pip install defense layer that constrains AI coding agents through enforced protocols; not prompts. Pre-flight checklists, HMAC-signed session audits, shell command boundaries, 35 drift checks, and a document type system that prevents docs from becoming lies. Works with any agent: Kimi, Claude, Copilot, Codex, custom.
Lightweight workflow guardrails for AI-assisted engineering.
Tamper-evident audit logging for OpenClaw and other local AI agents: hash-chained, locally verifiable, zero cloud dependency.
Pre-flight spend authorization for AI agents. Reserve-confirm lifecycle, concurrency-safe, append-only audit ledger.
Fast, local prompt-injection & jailbreak guardrail for LlamaIndex RAG — blocks attacks before retrieval. Powered by Bastion Prompt Protection (ONNX, ~5 ms, fully offline)
How to build a secure harness with sub-agents; Real Pi sub-agents, Cedar authorization on every spawn, tool call, and routing
Governance patterns for autonomous AI agents in regulated financial services — DEFCON state machine, Sovereign Veto, Audit Chain, EU AI Act mapping
MPC Elevation bridge for controlled privilege escalation
Counter-recon honeypot against agentic LLM attackers (PentestGPT/AutoGPT/LangChain). Six deploy modes: Flask drop-in, WSGI/ASGI middleware (Django/FastAPI), callable Python primitives, decoyshield CLI, edge configs (nginx/Caddy/Cloudflare), and Node.js npm package (Express/Fastify). Invisible-to-human, visible-to-LLM bait halts AI scans.
An agentic fairness and AI-risk analysis platform: bias detection for text, images, and datasets.
Dhi - Runtime Intelligence & Protection System. Sanskrit: Intellect | Perception. Kernel-space eBPF security monitoring for AI agents.
LLM Security & OWASP LLM Auditing Framework
Python SDK and MCP server for Polaxis - AI agent governance platform. Evaluate every tool call against policies in real time: block dangerous actions, enforce budgets, route to human approvals. Under 5ms.
Runtime control, security, and governance for AI agents — decide what an agent may do at the moment it acts, revoke that authority instantly, and keep tamper-evident evidence of every decision.
Governance-first framework for auditable AI agents: tamper-evident audit and cryptographic owner control (MAIC), persistent personality (HIM), agent runtime (NHE), pre/post-action verification and EU AI Act mapping, LLM-agnostic in TypeScript.
gradio region:us
gradio region:us
Python library that checks whether the next move in an agent workflow is allowed, and why. MIT licence. Copyright CONSOLIDARE CONTINUUM S.R.L.
Afterlogin: The Hunt × Helper Patrol — a cinematic identity & data-governance training game powered by real gpt-4o tool-calling agents + Foundry IQ grounding. Microsoft Agents League · Creative Apps.
Governance for AI agents: a deny-by-default policy layer for Claude Code and agent/skill builds. Every tool call is authorized, contained, and audited. Apache-2.0.
Dual-Auth provides AGBAC dual-subject Authorization for AI Agents and Humans using existing IAM Solutions (EntraID, Okta, Auth0, Keycloak)
AI Governance Infrastructure — open gateway demo. Self-hosted systemprompt.io gateway in a single Rust binary: policy, audit trails, and cost attribution on every AI request, with the Systemprompt Bridge desktop app for Windows and macOS connecting Claude Code and Claude Cowork.
Official GitHub Actions for Humanbound — adversarial security testing for AI agents in CI.
Making kimi --prompt safe to run unattended: nine findings on Kimi Code CLI permission behaviour, plus a fail-closed read-only Bash guard with a 125-case offline test suite.
AI Content Guard 2026: Adversarial Fact-Check Gate for Error-Free Code & Copy
RSCT Workflow Framework — memory + MCP governance that stops AI coding agents from skipping plans, bypassing authorization, or breaking conventions.
A modular governance framework for multi-agent AI teams
The trust control plane for enterprise AI agents — real-time policy verdicts in under 800ms and litigation-grade evidence of every decision. Pharos decides. Pharos proves.
Verifiable agent identity for A2A and MCP. Cryptographic claims, append-only transparency log, recovery-key revocation. Apache 2.0.
Whose Side Is Your Agent On? PrincipalBench — a multi-turn benchmark and post-training methods (prompt scaffold + per-token-KL distillation) for multi-party principal loyalty in LLM agents.
Inter-agent security firewall for multi-agent AI systems: runtime prompt-injection filtering, Ed25519 trust attestation, and capability enforcement for LangChain, LangGraph, CrewAI, and AutoGen. pip install inter-agent-guard
Daimon: The Personal AI Conscience for the Agentic Era — an open-source framework for building a conscience layer for personal AI agents
Official repo for of the ACL 2026 paper "Peering Behind the Shield: Guardrail Identification in Large Language Models"
A runtime assurance kernel for LLM-based systems: graded verdicts, hashed commitments, bit-identical replay.
Hard permission boundaries for Codex on Windows — with verification, backups, and recovery.
Permissions that narrow when one AI agent delegates: a sub-agent never holds more than its parent. In-process, with a hash-chained audit log you verify offline. Python and TypeScript, adapters for 19 frameworks and A2A.
Safe Multi-Agent Deployment Platform: pairwise safety analysis for autonomous agents that share a filesystem, clipboard, and OAuth scopes. Every verdict is graded on a four-rung evidence ladder, from docs-only to executed-in-sandbox.
Deterministic AST linter for multi-agent systems. Finds unattenuated agent delegation. No dependencies, no network, never executes the code it reads.
An undo layer for AI agents. Records what an agent changed through MCP, puts it back, and refuses to let it do the unforgivable unasked.
MCP server for Agent Trust Stack — cryptographic provenance and reputation scoring for autonomous AI agents. Install: pip install agent-trust-stack-mcp
[ASE 2026] ProvenanceGuard is a provenance-based runtime guardrail for LLM agent safety that detects misaligned tool calls before execution and keeps agent actions aligned with user intent.
Open-source security & behavior review platform for AI agents. Defense-in-depth: pattern scanning, task-aware semantic review, and deterministic capability sandboxing.
docker llm-security sql firewall ollama gradio region:us
Runtime guardrail layer for AI agents: detect loops, retries, cost explosions and lack of progress, and intervene during execution.
DSH 插件:把权限审批和模型提问改成大白话;审批按风险分级并列出会改动的文件路径。Plain-language approvals and question prompts for DeepSeek Harness.
Pre-install trust gate for coding agents. Check pip/npm before install; never runs the package.
LLM Security Platform Docs
Deterministic governance runtime plugin for OpenClaw.
Declarative specification for AI agents to publish identity, capabilities and operational boundaries before interaction.
The Execution Security Layer for the Agentic Era. Providing deterministic "Sudo" governance and audit logs for autonomous AI agents.
Developmental lifecycle architecture for AI agents: childhood stages, mortality, sealed wisdom distillation, soul persistence, and relational witnessing. From mountain night shifts to open garden.
The PSA Control Standard defines governance controls for deployed agentic systems.
Smallest outsider-readable proof path through AgentGate + MCP Firewall: governed write_file, independent on-disk verification, and inspectable audit artifacts.
Research constitution and specification framework for comparing independently governed AI judgments without transferring authority.
Runtime defense for AI agents. 24 inline defenses, 3 output scanners, MCP server, framework adapters.
Vienna OS — The execution control layer for autonomous AI systems. Cryptographic warrants, risk tiering, policy enforcement, and immutable audit trails.
Middleware library for AI accountability in LLM-based multi-agent systems — provenance logging, responsibility scoring, risk escalation
Multi-agent debate framework for enhancing LLM safety through red-teaming prompts, feedback-driven learning, long-term memory, and diverse structured debate strategies.
Safety kernel for agentic AI — multi-model consensus, deterministic verification, SHA-256 audit trails, checkpoint/rollback, self-healing. Rust + Tokio.
Deterministic Runtime for AI agent tool calls. Intercept payments and emails before execution
Don't let your agent delete the database. Hybrid static + LLM permission classifier for Claude Code, OpenClaw and Cursor.
玄武 Genbu — AI 防禦與保護模組。防止 prompt injection、記憶污染、鏈式攻擊與跨實例污染。基於 LDRIT 設計。
🎭 Portable non-human identity specification (persona.json metarepo)
CoreBlow trust and safety resources.
See what your AI coding agent actually did. Local audit of Claude Code session transcripts: leaked secrets, risky shell, unsafe edits, hook bypasses. Cursor + Windsurf adapters on the roadmap.
Deterministic stateful governance for AI agents and agentic systems