AI guardrails & agent security
3,625 agents · most starred first
AI guardrails & agent security, page 13 of 37
Open-source, 100% reproducible AI Agent Runtime Security Benchmark & Sandbox Environment (RFC-010 Draft Protocol).
A semantic type system for validating LLM output intent — Pydantic for meaning.
Synchronous JS/TS gate that catches destructive AI-agent tool calls before they execute. AST-level SQL parsing, SSRF defense, shell metacharacters, path traversal.
Eticas AI Risk Taxonomy: Open, machine-readable AI risk taxonomy mapped to the EU AI Act, NIST AI RMF, ISO/IEC 42001, and 15+ other frameworks.
One agent proves a capability is safe. Every other agent -> Codex, Cursor, Claude Code, reuses it instantly, without rebuilding or re-verifying.
Agent Receipts / Obsigna plugin for OpenClaw — cryptographically signed audit trail for agent actions
Agent-first secrets, zero .env, zero servers. Encrypted per-env dotenv files committed to git via sops+age. Go + Zig CLI.
LEASH-8: an 8-domain control model for AI agents with delegated authority. Scorecard, approval-design checklist, plan-vs-authorize pattern. Patterns from a live production agent operation, sanitized. Free, MIT.
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.
Open-source tools for bounded AI-agent handoffs: explicit file packets, evidence receipts, controller-side verification, and a native Hermes plugin.
Action Commit Safety for AI agents: the approved tool call is the one that executes, and uncertain side effects are never blindly retried. Framework-agnostic Python runtime with durable approvals, idempotent execution, and signed audit.
Open Agent Security Fabric — red-team agentic AI, MCP firewall SDK, action oracles & SARIF CI gates. Assure, Enforce & Govern. Python + TypeScript. OWASP LLM/Agentic/MCP Top 10.
Defense-in-depth against prompt injection in tool-using LLM agents. Semantic drift detection catches synonym-evading attacks that bypass regex. Self-play red-team, calibrated scoring, ablation study.
Own your agent security: a deterministic, offline firewall for AI agent tool calls. Same call, same verdict, no model in the path.
Zero-trust security runtime for AI agents with deterministic policy enforcement, guarded execution, approvals, redaction, MCP protection, and tamper-evident auditing.
Message notification router: Haiku safety gate + Opus router with code-enforced guardrails. #94 of 1,983, HackerRank Orchestrate 2026.
AgentInterdict — runtime enforcement for autonomous AI agents. Trust context. Verify authority. Interdict unsafe action.
Compiler-style pre-execution safety analysis for Harness shell actions.
Find and scrub secrets from AI coding agent session logs and configs: Claude Code, Codex, Gemini CLI, aider, Cursor and more
A modular, extensible toolkit for red teaming large language models and NLP systems — covering adversarial text attacks (character, word, sentence, semantic), jailbreak evaluation via JailbreakBench, and prompt injection — with pluggable model targets, automated judges, and clean reporting. Built for researchers and AI safety practitioners.
Bounded Agents: Delegation Security for Multi-Agent AI Systems - Reference Implementation and Evaluation Artifact
Prevent Copilot from reading your secrets using .pilotignore. Works like .gitignore but for GitHub Copilot.
Tool-call guardrails for Claude Code derived from cc-logger telemetry: nudge recoverable mistakes, block irreversible ones. The guard stage of callusguard; still works standalone.
AI Agent Scanner - A standalone AI Agent scanner that can discover agent repos, configuration, system prompts, classify risk and produce compliance-ready reports
Security Agent Benchmarking and Evaluation Framework
Cross-Vendor AI Safety Validation Report GigaChat Prompt Jailbreak — Transferred to Yandex Alice AI
Semantic permissions for AI coding agents. Intenter remembers what commands actually do, not just what they’re called.
Security enforcement for AI coding agents. Catch and block insecure code.
Provenance-tracking firewall that stops LLM agent prompt injection (OWASP LLM01) via token-level taint analysis — a zero-trust, deny-by-default guard on privileged tool calls.
Sandboxed security scanning for AI skills and MCP servers — precise findings, heatmap posture, fail-closed agent guard. Multi-scanner security platform for AI skills and MCP servers, scan what the agents install!
Open-source runtime security for tool-using AI agents. Control risky actions, enforce policies, and keep audit trails for MCP and agent workflows.
🔍 面向 AI Agent 的防篡改审计溯源系统 | Tamper-evident audit trail for AI Agents — generates hash-chained JSONL logs to answer: what did the agent do, how did it decide, and was anything altered? Framework-agnostic core + OpenClaw plugin.
Evidence-bound assurance for agentic AI — governance, provenance, authorization, audit, and fail-closed lifecycle controls.
Python contracts for AI-agent tool results. Block empty, partial, stale and unverified JSON, HTTP, MCP and stdio observations.
Open-source identity and authorization layer for AI agents. Issue verifiable credentials, manage decentralized identifiers, and enforce scoped permissions — self-hostable, standards-compliant, and built for autonomous AI systems.
OS-level security supervisor for AI coding agents. Gates security-relevant syscalls on Linux and pauses ambiguous actions for human review.
Autonomous Security & Zero-Trust Governance Guardian for Enterprise AI Agent Fleets powered by Gemini & Google Cloud. Built for #AllThingsAgenticHackathon.
Lightweight prompt injection detector for LLM, RAG and AI agent applications using Sentence Transformers + scikit-learn.
R-CAGE: Recovery-conditioned agent governance for zero-trust, ransomware recovery, software supply-chain security, and crypto-agile agentic defense.
Sealed-lab research harness for emergent multi-agent (LLM) coordination and containment — replicates the dynamics of the 2026 OpenAI–Hugging Face incident against self-hosted, intentionally-vulnerable targets. Air-gapped by design. Defensive / AI-safety only.
OS keychain secrets for AI coding agents, over MCP. Native vault storage (macOS Keychain, Linux Secret Service, Windows Credential Vault), policy-governed access, audit log — CLI + 44 MCP tools.
Mission-bounded authorization mesh for humans, workloads, and AI agents — deterministic policy enforcement across RBAC, ABAC, ReBAC, OAuth, Kubernetes, and agentic tool calls.
Secure your AI agents with policy-based authorization. Wraps browser-use, Playwright, LangChain, and PydanticAI with pre-action guardrails and post-execution verification.
Capgate — runtime authority boundary for autonomous AI agents: the agent proposes, the harness authorizes, the executor acts. Deny-by-default policy, non-transitive delegation, tamper-evident audit, MCP gateway, containerised isolation.
local first tool to Stop rogue AI agents from deleting production. Runtime security layer that intercepts tool calls, blocks dangerous commands, and audits everything.
Signed, hash-addressed decision records for AI agents, with reasons you can re-run
Fast System-1 Decision, Arbitration & Safety Engine for Autonomous AI Agents (Powered by TypeSafe Jev)
SENTINEL: an authority-attenuating immune architecture for AI agents. Detect and contain compromised agents without giving the defenders unrestricted authority.
Find and scrub secrets from AI coding agent session logs and configs: Claude Code, Codex, Gemini CLI, aider, Cursor and more
Accountability layer for AI agents: passport, mandate, action gateway, tamper-evident journal — and undo
Deterministic action-governance kernel for LLM-driven systems with fail-closed execution, signed approvals, and verifiable audit chains.
An open protocol for approving AI agent tool calls before they run. Specification, OpenAPI contract, Go adapters and CLI.
9 humanity skills for AI agents — crisis detection, WCAG audits, empathy, cultural sensitivity, and more. MCP + npm.
Deterministic cryptographic runtime ledger, flight recorder, and zero-trust security perimeter for autonomous AI agent swarms. Built in Rust.
Envelope-encrypted team secrets your AI agent can use but never read. Zero dependencies, zero servers — your git repo is the backend.
Catch fabricated RAG citations before they reach the user: a zero-token verbatim gate + a burden-of-proof judge, framework-agnostic.
Permission firewall, auto-approve hooks & command guardrails for 24 AI coding agents — Claude Code, Codex, Cursor, OpenCode, Antigravity, Copilot, Cline, Kiro & more.
Provenance-aware authorisation for LLM agent tool calls. Denies a call when a consequential argument came from untrusted content. Stdlib only.
Kernel-level read gate for AI agents — protected file opens are denied or routed through off-band consent before bytes reach the agent. Linux (fanotify) + macOS (Endpoint Security).
The trust layer for AI agents.
AI agent desktop workspace for independent developers and creators: task-scoped permissions, prompt-injection checks on your own tasks, disclosure review, clean retries and revocable project memory. Bring your own model.
A local guard and egress proxy for always-on AI agents: key vault, prompt-injection scanner, leak blocker, hash-chained audit log, kill switch. Zero dependencies.
Records what AI coding agents like Claude Code and Codex do on your Mac: commands, files, secrets and network. CLI and menu bar app.
Sub-100ms security gate for AI agent tool calls, powered by TypeSafe's Jev (System-1) decision model. Single-request Choice/Noul/Score evaluation, dual-factor blocking matrix, calibrated-confidence routing, fail-closed parsing, zero-config local fallback. LangChain-ready.
Open, bilingual (AR/EN) AI-governance starter kit for GCC organisations: register AI use cases, get an explainable risk tier, export controls. Browser-only.
RAG retrieval-time poisoning detector — demand-recency discrimination, canary injection, provenance attestation; fully offline, SARIF+MD reports
Self-hosted AI orchestration with policy, approvals, PII controls, audit evidence, retrieval, and governed MCP tools
Ironheights: The Immune System for AI Agents. Protects any AI bot that installs it: scans agent skills (OpenClaw, Claude, Grok Bot, Codex, Cursor, MCP) A-F, threat-model PR checks, MCP server, drift guard, posture report. Local, no telemetry.
Open-source, self-hosted agent platform for banks and fintechs, declarative agents, banking guardrails, CI-gated evals, tamper-evident audit. Beta.
Content safety + governance for AI agents in one import. IBM Granite Guardian risk classification, NVIDIA NemoGuard content moderation, autonomous-send caps, and a tamper-evident audit ledger — over the three.ws Trust Layer.
stAirCase - The enforcement gate between AI agent plans and your codebase: approvals bound to exact bytes, a signed audit trail
PostgreSQL/MySQL wire-protocol proxy for database guardrails for AI agents: stops destructive SQL before it reaches your database, no code changes.
Learn to inspect external sources before AI agents use them: inert exercises, static risk receipts, and clear boundaries before execution.
Fine-tune, calibrate and gate Laya (open-weights System One decision model) on your own decisions, then guard production agents with it
[ACL 2026 Findings] "LeakDojo: Decoding the Leakage Threats of RAG Systems"
Self-hosted governance for AI agents: MCP & LLM gateway, signed policy enforced by an egress proxy and an eBPF sensor, agent-held Ed25519 + ML-DSA-65 keys and an offline-verifiable audit log. Even the server can't forge what an agent did.
Service-signed encrypted receipts for AI agent actions. https://sello.build/
Byte-level conformance vectors for Agent Passport System implementations. JCS canonicalization, delegation, decision receipts, instruction provenance. Bring your canonicalizer, verify against the reference.
Your AI agent ships hallucinated APIs and secrets. aiZee blocks them — policy, budget, guardrails and audit enforced on the action path for Cursor, Claude Code, Copilot, Windsurf, and MCP.
ABOM: a portable Authorization Bill of Materials schema and reference toolkit for human, workload, service, and AI-agent access
A local-first encrypted secrets vault for developers. API keys, credentials, commands and links — accessible with a hotkey.
I just found new way to jailbreak ChatGPT thast can do anything you asked it.
Reproducible audit of country-level government-ID support in OpenAI's Trusted Access for Cyber verification flow.
Scan AI coding-agent config files (CLAUDE.md, AGENTS.md, .cursor/rules, MCP configs) for hidden prompt-injection attacks.
Claude Code plugin: read-only reader agents plus deny-only hooks that keep web pages, documents and strangers' code from steering agents that hold a shell.
A lightweight, zero-cost Python proxy designed to intercept, audit, and sanitize developer prompts before they reach external LLM APIs. Implements real-time tokenomics, FinOps cost controls, and automated PII/secret scanning for secure AI engineering.
Admin role management for Microsoft Entra and Azure. See who holds admin roles in Entra and Azure RBAC, from users and guests to apps and AI agents, and change Entra PIM in bulk. Open source, runs in your browser.
AI that thinks in layers, not just tokens. Intent-based processing through 15 dimensional analysis. Anti-manipulation through understanding context. Built with Rust + Go.
AEGIS OS™ - Secure AI Agent Platform - 12 security layers, 5 AI backends, RBAC, TLS, MCP/A2A. Built in Rust.
A1 – Know Your Agent Verifiable Authorization for Multi-Agent AI
Sovereign Intelligence System - Equilibrium-native reasoning for OpenClaw Safety
jingu-trust-gate – deterministic admission layer that blocks LLM hallucinations from becoming system state
OS-level runtime security for AI agents. Tamper-proof monitoring, behavioral detection, and audit trails.
Governed context projection for AI work: receipts for what AI tools are allowed to see.
Make sure AI agents actually do what they claim, no more fake successes.
AgentShield is a middleware implementation of Zero Shot Embedding Drift Detection method for detecting and deflecting prompt injection attempts.
Passport & visa system for AI agent credentials. Detect, classify, and govern API keys, tokens, and secrets across MCP, OpenAI, Anthropic, LangChain, and more. Encrypted vault, policy engine, audit trails, delegation chains. Drop-in MCP server replaces plaintext credential storage.
Security enforcement plugin for Claude Code. Blocks dangerous commands, audits every tool call, detects prompt injection.