Security & AI safety

AI guardrails & agent security

3,625 agents · most starred first

AI guardrails & agent security, page 16 of 37

ivycheck-python-sdk★ 2

Python SDK for IvyCheck

garl★ 2

Cryptographic verification for AI agent actions — ECDSA-secp256k1 + RFC 6979 signed Action Receipts (v0.1), multi-dimensional Trust Vector, capability tokens (JWT-shaped + Biscuit-style attenuation), UETA §10(b) undo. 29 MCP tools. A2A v1.0 + ERC-8004 format compatible. Receipt-batch Merkle roots anchored on Base mainnet. Starting with code.

AgentPass★ 2

Runtime control and observability for AI agent actions: policy, approvals, durable boundary evidence, outcome assurance, idempotency, and OpenTelemetry correlation.

scbkr-local-responsibility-model★ 2

Local AI responsibility-chain control layer with owner-signed Workbench, SCBKR S/C/B/K/R grammar, Data Center, four-store evidence reuse, FastAPI, React, and Tauri desktop runtime.

belief-state-runtime★ 2

Epistemic reasoning engine for AI Agents — assess claim credibility with calibrated confidence. Includes a ready-to-use .skill package for agent frameworks. 87%+ accuracy with evidence, 85%+ abstain without.

witness-stand★ 2

Adversarial AI system to test and improve reliability under real-world pressure

Neuroverseos-governance★ 2

Deterministic governance engine for AI agents. Enforce rules defined in .md governance files across AI systems.

text-graphics-agent★ 2

A security-first agent platform — Models Propose, Records Decide. 17 deterministic constraints gate every LLM proposal before it becomes state.

agent-firewall★ 2

A dry-run firewall for coding agents — approve & audit tool/MCP/shell calls before they hit the real world. By Martello Systems.

agentguard★ 2

agentguard

cavra★ 2

CAVRA - Controlled Agentic Verification & Runtime Authority. Before the agent acts, CAVRA decides. CAVRA is a runtime governance and authority layer for AI coding agents.

PACT★ 2

PACT is a runtime security layer for AI agents.

sessionguard★ 2

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.

ayiru★ 2

Verified, machine-readable safety verdicts for AI agents running shell commands. Append-only audit log, six verification levels, deterministic risk engine. MIT.

finance-chatbot-with-guardrails★ 2

gradio region:us

ai-human-text-detector-llm★ 2

streamlit region:us

ora-grounding★ 2

Post-response grounding check + cross-family adversarial review for LLM chat agents

agent-leash★ 2

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.

leadline★ 2

Evidence router & policy engine for coding agents — enforces source choice and proof-of-use through Claude Code hooks, including routes to MCP tools. Local-first; no LLM in the hook loop.

headroom★ 2

More human oversight can make an AI agent less safe. Headroom is a human-in-the-loop firewall for coding agents that measures when to trust the human: oversight as resource allocation, not just classification.

action-graded-severity★ 2

Action-graded severity scoring (L0-L6) for tool-using AI agents, computed from red-team execution traces.

production-llm-agent-checklist★ 2

Security & runtime checklist for moving LLM agents from demo to production: HITL gates, idempotency, prompt injection boundaries, tool risk classification (CC BY 4.0).

mcp-capability-ledger★ 2

Infer least-privilege MCP tool policies from agent traces

valta-audit-chain★ 2

Upstream pre-call budget gateway and tamper-evident SHA-256 audit ledger for autonomous AI agents (CrewAI, LangChain, OpenAI).

maqam★ 2

Give AI agents hands without handing them the keys: exact one-use approvals and verifiable receipts for registered actions.

writ★ 2

Keep your agent and workflow. Add visibility first, limits when needed, and enforceable contracts for high-risk work.

genesis-os-blueprint★ 2

Architecture blueprint for accountable AI agents — a cognitive kernel, fail-closed policy hooks, swappable model providers, and evidence-graded outcomes.

Invart★ 2

Local-first control plane for governing AI agent runtime behavior, from preflight risk checks to live mediation and post-run audit.

inkie-auto-mode★ 2

Don't let your agent delete the database. Hybrid static + LLM permission classifier for Claude Code, OpenClaw and Cursor.

guard-bands★ 2

HMAC-signed boundaries that keep untrusted LLM content as data, not authority

openclaw-human-gate★ 2

Manual Verification Protection + AI Selector Compatible with OpenClaw . This is a complete TypeScript plugin project—a true implementation of the `before_tool_call` hook that can intercept tool calls, follow OpenClaw’s built-in approval workflow, and features an approval window mechanism.

quesen★ 2

Quesen — Deterministic AI decision engine for autonomous-agent risk evaluation. Native MCP server + Agent Settlement Protocol (ASP/1.0). Zero-LLM. Same inputs → same output. Live: web-production-aa5ba.up.railway.app/mcp

faramir★ 2

A secrets broker for local AI coding agents: it runs the commands that need credentials and keeps the values out of the agent's context.

AuditronClaw★ 2

透明可审计的本地智能体终端:对开源 agent 框架做安全审计与加固(RCE/沙盒绕过/注入),并以 50 条注入基准 × 37 条能力基准双维度量安全边界

gate.cat★ 2

Block irreversible AI-agent actions before they run — deterministic, fail-closed action veto for Claude Code, Codex, Cursor, and other tool-using agents.

agent-approval-gate★ 2

The layer a sandbox doesn't give you: a PreToolUse hook that blocks unapproved spending, account creation and fund movement before the call runs, queues them for human review, and proves it actually ran -- including whether your harness fired it for subagent tool calls. Plus a spend ceiling and loop detector.

cloudguard★ 2

Reliability and safety harness for LLM cloud agents. Tool-selection accuracy vs. tool count, blast-radius guardrails, and prompt-injection red-teaming against a real AWS mock. Every number is a committed JSON artifact.

openspine★ 2

A safety layer for AI agents; it decides what your agent is allowed to do - and proves it.

toolcall-contract★ 2

Deterministic two-layer validator for LLM tool calls: stdlib structural checker in 25,000-case differential agreement with pinned jsonschema, plus the semantic contract layer where real failures live - qwen2.5-7B: 0 schema violations, 7 semantic, replayed in CI from committed data.

ActionGuard★ 2

A local, user-controlled safety boundary for AI-powered automation.

touchid-agent-guard★ 2

Touch ID confirmation for AI coding agents on macOS. Pair with sandboxes for defense-in-depth against prompt injection.

sosa★ 2

SOSA™ — Supervised Orchestrated Secured Agents. A governance framework for production-grade autonomous AI operations.

guardian★ 2

A security layer for AI agents. Stops prompt injection, tool tampering, and runaway behavior before they happen. Open-source guardrails for AI agents — auditable, self-hosted, and works with any framework. Deterministic security sidecar for LLM agent frameworks. Drop-in protection against prompt injection, tool poisoning, and capability abuse.

secchecker★ 2

Static security scanner and analyser for AI agents, MCP tools, and LLM applications - OWASP LLM Top 10, and infrastructure checks. Zero runtime dependencies.

agent-browser-plugin-proton-pass★ 2

Proton Pass credential provider for agent-browser — resolves logins so the password never enters the agent's transcript, argv, or shell history. Zero runtime dependencies.

unanswered-approval★ 2

Runnable, offline reproductions of one defect class in three agent frameworks: a value that is neither an explicit yes nor an explicit no reaches the execute path of a human-approval gate.

IX-BlackFox-Cognition★ 2

Governed cognition substrate for AI engineering agents: mission envelopes, evidence-bound belief/plan graphs, quarantined memory, model-role review, sentinel checks, and BlackFox handoffs under human authority.

agentguard★ 2

Runtime security layer for AI agents — inspect, control, and audit every tool call. Trust-aware data flow, 3-layer intent consistency detection, Merkle audit trail. Drop-in support for LangChain, CrewAI, AutoGen, OpenClaw, MCP, Dify, AutoGPT, n8n.

leash★ 2

Keep your AI agents on a leash. API-layer policy engine for authorization, audit, and security features which includes deny-by-default, simple YAML rules, and tamper-evident logs.

agentguard★ 2

Welcome to cerbere-agent guard, originally just agent guard, then rebranded after the mythical hound of the underworld. Our mission? Guarding your AI agents, keeping them in line, and blocking unauthorized actions. Think of it as the ultimate guard dog for AI security and guardrails. 🐕🛡️

nfi-tools★ 2

CLI tool and MCP server for handling sensitive information without exposing values to AI assistants

agent-trace-witness★ 2

External witness for autonomous multi-agent AI systems: signed readiness seal, choke-point capture, and PROV-DM causal graphs for post-incident reconstruction

voice-ai-governance★ 2

Compliance enforcement middleware for voice AI pipelines: warm transfer state management, PII scrubbing, confidence-gated escalation, and HIPAA/FERPA/EU AI Act enforcement for Pipecat, LiveKit, and Twilio deployments.

agent-containment-replay★ 2

Replayable agent trajectories reconstructed from the July 2026 frontier-lab evaluation-containment failures, plus benign controls for measuring false-positive cost.

ttp-protocol★ 2

See what your AI agents can cause, route them safely, and prove what happened. Open-source execution authority for AI agents — consequence preview, trust routing, signed receipts.

AgentGuard★ 2

As autonomous AI agents are increasingly deployed across enterprise workflows, their tool-calling capabilities (e.g., executing shell commands, sending emails, or modifying databases) expose critical security vulnerabilities. Prompt injections or reasoning loops can lead to catastrophic unauthorized actions—such as dropping database tables or leaki

stepwarden★ 2

Every tool call your agent makes, checked before it runs. A Claude Code plugin that uses TypeSafe AI's Jev to verify each pending tool call against the session plan, then allows it, asks you, or blocks it. Proof of concept

agentstamp★ 2

🔏 Trust verification for AI agents — Ed25519 stamps, trust scoring (0-100), x402 micropayments, 14 MCP tools. Free tier.

ai-agent-incident-explorer★ 2

static ai-safety ai-security autonomous-agents incident-database sandbox-escape benchmark agent-containment

SODA-Agent-Safety-Judge★ 2

safetensors qwen3 safety judge agents tool-use

common-sense-skill★ 2

One file. 9 questions. Stops AI agents from doing the obviously-wrong thing. MIT. Works with Claude Code, Codex, Cursor, OpenClaw.

agent-handrail★ 2

一个架在任意 agent harness 之上的**增量安全/可用扶手层**。不替换原 agent,只加护栏。

qed-proof-js★ 2

QED Proof TypeScript SDK and CLI (@qed-proof/sdk, @qed-proof/cli): verify AI agents' claimed work and check signed receipts offline from Node.js.

panopticon-protocol-v3★ 2

A counter-espionage RL environment for scalable AI oversight, deception detection, sleeper-agent research, and reproducible security-first training.

Cognitive-Friction-Toolkit★ 2

No Promises Without Proof. A dependency-free discipline that forces agents to pair every success claim with evidence and a proof state (OBSERVED/INFERRED/UNVERIFIED/BLOCKED), then derives the completion verdict deterministically.

agent-rule-gaming★ 2

Experiments in a small RL environment: do LLM agents game the rules when the answer is within reach, and can you tell them not to? Claude Opus 5, Qwen 3.8 27B, abliterated Qwen.

provenanceguard★ 2

Source-aware factuality verification for MCP-based LLM agents — paper artifact

Adapti-guard★ 2

Empirical case study of measurement validity in LLM-agent prompt-injection defense evaluation: how success endpoints, labeling rules and reporting units change conclusions on the same traces.

hitl-guardrail★ 2

A human-in-the-loop policy guardrail for LLM / AI-agent actions — approve, escalate, or reject with an auditable reason and risk score. Zero deps, typed, framework-agnostic (CrewAI/LangGraph).

vulnify-sdk-python★ 2

Python SDK and CLI for Vulnify: check an AI agent action and get ALLOW, REVIEW or BLOCK before it runs. Adapters for CrewAI and LangGraph.

strathon★ 2

The open-source AI agent firewall — write a CEL policy, block the tool call before it executes. In-process, sub-millisecond. OWASP + EU AI Act mapped.

devboost-utility-api★ 2

Next-gen utility API for AI agents & developers: Prompt Injection Guardrails, LLM context sanitization, web scraping, and data validation.

votumsecuris★ 2

Local-first governance layer for AI agents. Policy engine, audit trail, rollback, cost control. All offline.

AgentGuard★ 2

As autonomous AI agents are increasingly deployed across enterprise workflows, their tool-calling capabilities (e.g., executing shell commands, sending emails, or modifying databases) expose critical security vulnerabilities. Prompt injections or reasoning loops can lead to catastrophic unauthorized actions—such as dropping database tables or leaki

safetydrift★ 2

First open-source implementation of the SafetyDrift framework (arXiv:2603.27148) — predicts when individually safe AI agent actions compound into a safety violation using Markov chain analysis. Works as an MCP middleware.

tempus-ddb★ 2

Fail-closed B2A security gate for autonomous agent actions with signed policies and verifiable receipts.

GATE★ 2

Zero-trust runtime security for LLM agents: validate every tool call before execution in LangChain, LangGraph, OpenAI Agents SDK, and CrewAI.

derbent★ 2

One guarded pass for all your coding agents. A derbent was a guarded post on an Turkish History mountain pass: its keepers decided who went through and kept a record of everyone who did. Derbent does the same for your coding agents' tool calls.

tempus-ddb★ 2
agent-cost-bench★ 2

Reproducible denial-of-wallet (DoW) benchmark for tool-using LLM agents, scored on provider-billed tokens/USD.

miseguard★ 2

Deterministic runtime circuit breaker and MCP proxy for autonomous coding agents.

Secure-AI-Agents-Suite★ 1

gradio region:us

Co-Pilot_Audit_Gym★ 1

docker openenv reinforcement-learning ai-safety governance compliance multi-agent rl-environment

llmscanner★ 1

streamlit region:us

huolongguo10-LLM_detect★ 1

gradio region:us

LLM_Detector★ 1

streamlit region:us

togethercomputer-GPT-JT-Moderation-6B★ 1

gradio region:us

HarmBench-Llama-2-13b-cls-multimodal-behaviors★ 1

transformers safetensors llama text-generation conversational arxiv:2402.04249

genesis-accord-standard★ 1

Open standard where an AI agent certifies its human operator. 8 evidence-backed dimensions. v4.0.

AIActGuard★ 1
aurascape

Aurascape review: compare features, pricing, use cases, access model, and alternatives for this AI Security agent in 2026. Keep secrets in, keep threats out. Allow AI and unleash productivity.

velatir

Velatir review: compare features, pricing, use cases, access model, and alternatives for this Authentication Agents agent in 2026. Rated 5/5 from 1 verified review. Where Innovation Meets Human Oversight

Aurascape

Keep your company's AI use safe and compliant.

Superagent

Keeps your AI apps safe and compliant, blocking threats in real time.

safety-agent

An open-source SDK for AI agent safety

agenshield

AgenShield — AI Agent Security Platform

@attestplane/attestplane

Apache-2.0 attestation and audit substrate for AI agent evidence chains.

ecc-agentshield

Security auditor for AI agent configurations. Scans Claude Code setups for vulnerabilities, misconfigs, and injection risks.

veto-sdk

Policy runtime for AI agent tool calls

governance-sdk

AI Agent Governance for TypeScript — policy enforcement, scoring, compliance, and audit for AI agents

More in Security & AI safety