convez.ai
AgenteSegurançaIntermediário
Secure-Vector

SecureVector AI Threat Monitor

Secure-Vector/securevector-ai-threat-monitor

Auditoria e governança em tempo real para chamadas de agentes IA

Ver no GitHub
BomPluginScore · de 0 a 100

Agente de segurança e observabilidade para IA que auditoria cada chamada de ferramenta, enforce políticas de allow/block, detecta injeção de prompt e vazamento de dados. Funciona local-first com Claude Code, Copilot CLI, LangChain e principais LLM APIs.

Instalação

$npx convez-ai Secure-Vector/securevector-ai-threat-monitor

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
15
Forks
1
Acompanhando
2
Saúde
62%
Último commit
há 1 dia
Versão
v6.1.0
Issues abertas
2
Contribuidores
1

Casos de uso

  • Monitorar e bloquear chamadas de ferramentas potencialmente maliciosas
  • Detectar tentativas de injeção de prompt e vazamento de dados
  • Auditar histórico completo de execução de agentes IA
  • Enforçar políticas de segurança em agentes LangChain, CrewAI ou custom
  • Governança local sem telemetria ou dados saindo do servidor

Veredito

Ferramenta madura e bem-documentada que preenche gap real de segurança em pipelines de agentes IA. Abordagem local-first é acertada; integração com Claude Code e LangChain é genuína, não cosmética.

README

A tradução para o português deste README ainda está na fila.

License PyPI Python npm CLI Downloads/month Downloads total Discord

Start here

Pick the one that describes you.

I use Claude Code, Codex, Copilot CLI or OpenCode

No code changes. Install the app, then install the plugin for your agent.

pip install "securevector-ai-monitor[app]" && securevector-app
npx @securevector/cli                              # or from npm (needs Python 3.10+)

In the app, click Connect Agents, pick your agent, click Install Plugin, then restart the agent (in Claude Code, run /reload-plugins). From then on every tool call it makes is checked and logged. To run sessions inside the app, click Agents, then + Launch (full terminal on macOS and Linux; output only on Windows for now), or from a terminal:

sv-monitor session launch claude-code ~/project    # comes with the app; npm: securevector monitor session ...

More in the Agent Sessions guide.

I build agents in Python

LangGraph, LangChain or CrewAI: use the SDK for your framework (it brings the app with it). It checks every tool call before it runs.

# pip install securevector-sdk-langgraph
from langchain.agents import create_agent
from securevector_sdk_langgraph import secure_middleware, cost_tracking_middleware

agent = create_agent(model, tools, middleware=[
    secure_middleware(mode="enforce"),   # checks every tool call before it runs
    cost_tracking_middleware(),          # records model tokens and cost
])

Same idea for LangChain and CrewAI; create_react_agent and raw StateGraph patterns are in the LangGraph SDK guide. The framework SDKs record model cost themselves, so do not add instrument() below as well (why).

Plain OpenAI or Anthropic code: the app package includes a tracer and a tool guard. Guide

from openai import OpenAI            # or: from anthropic import Anthropic
from securevector import guard, instrument

instrument()                         # records every OpenAI and Anthropic call: tokens, cost, verdict
client = OpenAI()

@guard(tool_id="orders.lookup")      # checks this tool's input and output
def lookup_order(order_id): ...

with guard.session("ticket-8812"):   # groups one run on the Observability page
    ...                              # your agent loop: model calls and tool calls

I build agents in JavaScript or TypeScript

Run the app from npm. Needs Node 20+ and Python 3.10+ on your PATH; the launcher sets up the rest on first run. No Python? Use a binary installer instead.

npx @securevector/cli

That is all Claude Code, Codex and the other coding agents need; install their plugin as in the first path.

Optional: check the tools in agents you write yourself.

npm install @securevector/sdk
import { guard } from '@securevector/sdk';

const lookupOrder = guard(async ({ orderId }) => fetchOrder(orderId), { toolId: 'orders.lookup' });

Wrap each tool's function with guard() (with the Vercel AI SDK, the tool's execute), whichever framework you use: LangChain.js, the Vercel AI SDK, Mastra or plain Node. API and examples in the SDK repository.

Whichever path you took, open http://localhost:8741: every model call and tool call is there, allowed or blocked, with its cost and trace.

I'm evaluating it for my team

  • Where it runs and what leaves it: on each developer's machine, bound to 127.0.0.1, no telemetry. Outbound traffic only for the one-time Guardian model download, SIEM destinations you configure, and Cloud Connect if someone turns it on.
  • What the record shows: every tool call in a SHA-256 hash-chained log with a per-rule evidence ledger for blocked actions. The local hash chain shows tampering on this machine; the SIEM Forwarder gives you an off-host copy, and Device Identity ties events to a machine.
  • Who controls enforcement: locally, the developer, and every connector fails open, so an agent keeps working unchecked while the app is stopped. Central, read-only policies come from Cloud Sync.
  • Licence: Apache 2.0. Read the code, run it air-gapped, or self-host the engine. Full notes, including retention and what the app can see: Evaluating it for a team.

What you get

  • Run it here. Launch Claude Code, Codex, Copilot CLI or OpenCode from inside SecureVector and watch the terminal live, with every call and its verdict beside it. Sessions you started in your own terminal can be adopted onto the same board.
  • See it. One trace per agent session. Pick an agent on the left, read its whole run as a waterfall on the right. Live follow, replay, and a costliest-turn mark on every run.
  • Stop it. Allow, block or log-only per tool, enforced by default. A blocked call is refused straight away and becomes a request you can approve for fifteen minutes or an hour, after which the agent can retry. Blocking on detected threats in prompts and responses is opt-in.
  • Catch it. 108 rules covering the OWASP LLM Top 10, MITRE ATT&CK-mapped patterns and agent-attack chains, plus an optional offline ML model, applied while the agent is still running.
  • Know where it went. Every external host an agent reached, when it was first seen, and a policy that decides which ones it may reach.
  • Keep the record. Every tool call in a SHA-256 hash-chained log; the local chain shows tampering on this machine. Blocked actions get a per-rule evidence ledger.
  • Pay less. A local scan of your own transcripts shows why sessions cost what they did, with copyable fixes.
  • Keep it. Apache 2.0. No signup. Nothing leaves your machine unless you connect it.

Works with

You useHow to connect
Claude Code, Codex, GitHub Copilot CLI, OpenCodeConnect Agents in the app, pick yours, Install Plugin. Launch sessions from Agents if you like (guide)
Cursor, Antigravity, OpenClawConnect Agents, pick yours, Install Plugin (OpenClaw guide)
LangGraph, LangChain, CrewAI, HermesThe SDK for your framework: langgraph, langchain, crewai, hermes
Any other Python agent (OpenAI, Anthropic, Bedrock, Gemini, plain functions)@guard and instrument(), included in the app package
Any JavaScript or TypeScript agent (LangChain.js, Vercel AI SDK, Mastra, plain Node)@securevector/sdk
Anything that calls an OpenAI-compatible API (Ollama, Groq, n8n, Dify, any HTTP client)The LLM proxy: Connect Agents, Start Proxy, then set OPENAI_BASE_URL (or baseURL in Node) to it. Details
Claude DesktopMCP server
OpenTelemetrySend OTLP/HTTP JSON to http://127.0.0.1:8741/v1/traces. Tracing guide

If ports 8741 or 8742 are taken, pass --port. Prefer an installer? Windows, macOS and Linux builds need no Python.

Website · Getting Started · Verify your install · Discord · Screenshots

What's new in 6.1

  • Runs grouped by harness: Observability opens on Runs, grouped by harness with each session under it, running counts on every row, and a one-line summary (runs, sessions running, waiting, refused, median). A run's tool calls expand under the row; Open details keeps Session governance beside the trace.
  • Live grid: every running session as a tile with what it is doing now, waiting and blocked chips, tiles that open as a 2x2 or 3x3 matrix, and a running count on the tab. Sessions end on the harness's own end marker.
  • Agent Detection & Response: every governed session gets a score from 0 to 100 against what the same agent tool normally does in that folder, with the reasons in plain words. Two named rules act at call time: a credential file read followed by a network call, and outside content (web pages, MCP results) followed by a reach outward.
  • Approvals from the cloud (Professional and Enterprise): an enrolled device can let held calls be answered from Fleet Runs, off by default, with Deny always and the durations you allow. The device pulls the answer; nothing connects in.
  • Remote access mode (off by default): approve, deny and stop sessions from your own browser on another machine over HTTPS, with a short-lived token and re-sign-in for anything that lifts a block.
  • Guardian stays quiet: new findings show as an unread count and one short toast; urgent advice sits in the corner.
  • One-session approvals: approve a blocked host for 15 minutes, 1 hour or the rest of the session from the Approval inbox. Other sessions stay blocked.
  • Agent Config Trust: Integrations scans each agent tool's MCP servers, hooks and rules files, and shows changes as side-by-side cards to approve.
  • SecureVector MCP: check_policy tells the agent whether a tool call will be allowed before it makes it; session_burn reports tokens and estimated cost so far.
  • Fleet (Professional and Enterprise): Runs grouped by harness with inline expansion, run traces with governance, a Live grid across devices, held counts, rule matches and stopped sessions.

6.0 (Agent Sessions, Observability, npm install) and earlier releases are in the CHANGELOG.

Screenshots

More screenshots: Agent Map, Dashboard, Cost / Token Optimizer, Skill Scanner.

Install

MethodCommand
pip (Python 3.10+)pip install "securevector-ai-monitor[app]" then securevector-app
npm (Node 20+, Python 3.10+ on PATH)npx @securevector/cli, or npm install -g @securevector/cli then securevector
Installers (no Python).exe, .dmg, .AppImage, .deb or .rpm from the latest release, SHA256 checksums beside each file. Verify your install
Self-hostRun the engine in your own cloud with one Terraform module per provider: AWS, Azure, Google Cloud, Oracle Cloud

Only download installers from this repository. The npm launcher details, SDK-only and MCP extras, and updating are in the Installation guide.

Documentation

Contributing

git clone https://github.com/Secure-Vector/securevector-ai-threat-monitor.git
cd securevector-ai-threat-monitor
pip install -e ".[dev]"
pytest tests/ -v

Contributing Guidelines · Code of Conduct

Feedback

A wrong detection is the most useful thing you can send. Every detection in the app has a Report this rule link that opens a prefilled false-positive report with the rule id, version and platform, and nothing else. Anything else: open an issue, say hello on Discord, or email contact@securevector.io. If SecureVector is earning its place on your machine, a star helps others find it.

Cloud (optional, opt-in)

A separate cloud product handles MCP tool-permission policy sync across enrolled devices, per-org audit attribution, and per-device fleet slicing. It also adds AI Agent Governance, your agents' governance posture rolled into a single score across the fleet (app.securevector.io/governance), plus EU AI Act orientation that maps your action-layer logging and tamper-evident tool-call audit to the relevant obligations (governance/eu-ai-act); orientation only, not legal advice. Sign in for the fleet-wide view; the local install already gives you the single-device snapshot. Strictly additive: the local install above works standalone without it. Details: securevector.io.

Three ways to connect the app, all from Settings, Cloud Connect: pair your account (a short code you confirm in the browser, gives you rule sync and optional cloud analysis), enroll the device with a single-use token from the cloud app (puts the machine in Fleet, with runs, traces, the Live grid and held calls answered from the cloud), or paste a key you already have. The Connect to Cloud guide has the diagrams and a "which one do I need" table. To approve and stop sessions from another machine without the cloud, see Remote access mode.

License

Apache License 2.0, see LICENSE.

The starter SIEM dashboard templates under docs/siem/ (Splunk XML, Sentinel workbook, Datadog + Grafana JSON) are MIT-licensed; see docs/siem/LICENSE and docs/siem/NOTICE for trademark disclaimers.

SecureVector is a trademark of SecureVector. See NOTICE.


Get Started · Documentation · Discord · GitHub Issues · security@securevector.io

Ver o original no GitHub

Bom para

Relacionados