AICM — Open-Source Agent Integrity & Compromise Monitor

◇ What AICM Detects

Real Threats to AI Agent Systems

As AI agents gain autonomy — executing code, accessing databases, managing credentials — they become attack surfaces. AICM monitors for the threats that matter.

⚠

Skill-Injection Attacks

Detects when adversaries attempt to inject malicious instructions into agent tool chains, overriding intended behavior to execute unauthorized actions or extract sensitive data.

🔒

Credential Theft

Monitors for unauthorized access to API keys, tokens, database credentials, and secrets that agents handle during normal operations. Alerts on anomalous credential access patterns.

☢

Auto-Quarantine

When compromise is detected, AICM automatically isolates the affected agent — preventing lateral movement across your agent fleet before damage spreads.

👁

Prompt Manipulation

Identifies attempts to override system prompts, bypass safety guardrails, or manipulate agent reasoning through adversarial inputs — a growing attack vector in production systems.

⚙

Tool Abuse Detection

Monitors agent tool calls for anomalous patterns — unusual API endpoints, unexpected file system access, excessive data exfiltration, or calls that deviate from baseline behavior.

📈

Behavioral Drift

Tracks agent behavior over time to detect subtle drift — gradual changes in tool usage, response patterns, or decision-making that may indicate slow-burn compromise or model degradation.

◇ Architecture

Sensor → Server → Dashboard

┌─────────────────────────────────────────────────────────────────┐
│                      AICM Architecture                        │
└─────────────────────────────────────────────────────────────────┘

  ┌─────────────┐    ┌─────────────┐    ┌─────────────┐
  │  Agent #1    │    │  Agent #2    │    │  Agent #N    │
  │ ┌─────────┐ │    │ ┌─────────┐ │    │ ┌─────────┐ │
  │ │ Sensor  │ │    │ │ Sensor  │ │    │ │ Sensor  │ │
  │ └────┬────┘ │    │ └────┬────┘ │    │ └────┬────┘ │
  └──────┼──────┘    └──────┼──────┘    └──────┼──────┘
         │                  │                  │
         └──────────────────┼──────────────────┘
                            │ HTTPS / mTLS
                            ▼
  ┌───────────────────────────────────────────────────────────────┐
  │                     AICM Server                              │
  │  ┌──────────────┐  ┌───────────────┐  ┌──────────────────┐  │
  │  │ Policy Engine │  │ Telemetry Ingest│  │ Quarantine Mgr  │  │
  │  └──────────────┘  └───────────────┘  └──────────────────┘  │
  │  ┌────────────────────────────────────────────────────────┐  │
  │  │                FastAPI Backend                         │  │
  │  └────────────────────────────────────────────────────────┘  │
  └───────────────────────────────────────────────────────────────┘
                            │
                            ▼
  ┌───────────────────────────────────────────────────────────────┐
  │  React Dashboard  │  Agent Status  │  Alerts  │  Telemetry  │
  └───────────────────────────────────────────────────────────────┘

◇ Use Cases Across AI

Who AICM Is For

AICM is designed for any team deploying AI agents in production — especially in regulated or high-stakes environments.

◈

Financial Services

AI agents handling transactions, loan processing, or trading require security monitoring that meets regulatory standards. AICM provides the audit trail and quarantine capabilities compliance demands.

▣

Construction & Federal

Agents managing scheduling, RFIs, or USACE compliance on government contracts need verifiable security. AICM monitors agent behavior against defined policy baselines.

◉

Energy & Climate

Carbon tracking, ESG scoring, and grid management agents handling sensitive environmental data. AICM ensures data integrity and prevents unauthorized data exfiltration.

△

Aviation & Aerospace

Predictive maintenance and safety compliance agents where compromised output could affect flight safety. AICM enforces zero-trust monitoring at the agent level.

⬡

Defense & Logistics

Military and supply chain agents operating in adversarial environments. AICM's auto-quarantine prevents compromised agents from affecting operational decisions.

◇

Any Production Agent

If you're deploying AI agents that access APIs, databases, credentials, or make decisions — AICM gives you visibility into what your agents are doing and stops them when they shouldn't.

◇ Why AICM

The Strategic Advantage

Open Source Builds Trust

Full source visibility. No black boxes. Developers and security teams can audit every line. MIT licensed — use it anywhere, for anything, forever.

Compliance-Ready

Built for regulated environments. Telemetry logging, audit trails, and policy enforcement give you the documentation compliance officers and auditors require.

Framework Agnostic

Works with any AI agent framework — LangChain, AutoGPT, CrewAI, custom agents, or Centriv's own vertical agents. Drop in a sensor and you're monitoring.

Auto-Quarantine

Don't just alert — act. AICM can automatically isolate compromised agents, preventing lateral movement and damage before your team even sees the alert.

Enterprise Pathway

Free OSS core with an enterprise upgrade path: SIEM integration, Postgres telemetry, PagerDuty alerts, mTLS encryption, and compliance reporting dashboards.

Battle-Tested Methodology

Built by a team with 40+ years of military and aviation safety methodology. The same risk management rigor from Naval Aviation Safety Officer School now applied to AI agent security.

◇ Quick Start

Up and Running in 5 Minutes

Clone the Repository

Download AICM from GitHub.

git clone https://github.com/GReinhold-ai/aicm.git
cd aicm

Install Dependencies

AICM uses Python (FastAPI backend) and React (dashboard). Install the Python requirements first.

pip install -r requirements.txt

Start the AICM Server

Run the FastAPI server that receives telemetry from agent sensors and enforces policies.

# Development
uvicorn main:app --reload

# Production
gunicorn main:app -w 4 -k uvicorn.workers.UvicornWorker

Deploy a Sensor to Your Agent

Add the lightweight Python sensor to any AI agent. It monitors behavior and reports telemetry to the AICM server.

# Run once (for testing)
python agent_sensor.py --once

# Run as daemon (production)
python agent_sensor.py --server http://your-aicm-server:8000

Launch the Dashboard

View real-time agent status, alerts, and telemetry in the React dashboard. Or visit the live demo.

# Set up React dashboard
npm create vite@latest aicm-dashboard -- --template react
cp Dashboard.jsx aicm-dashboard/src/
cd aicm-dashboard && npm install && npm run dev

Or view the live dashboard: aicm-beta.vercel.app ↗

◇ Enterprise Edition (Coming Soon)

Production-Grade Agent Security

SIEM Integration

Forward agent telemetry to Splunk, Datadog, or any SIEM for unified security monitoring.

Postgres Telemetry Store

Persistent, queryable telemetry database for historical analysis, forensics, and compliance audits.

PagerDuty Alerts

Real-time alerting through PagerDuty, Slack, or webhooks when agents are compromised or quarantined.

mTLS Encryption

Mutual TLS between sensors and server — ensuring only authenticated agents can report telemetry.

Compliance Reporting

Pre-built reports for SOC 2, HIPAA, and federal compliance requirements for AI agent deployments.

Multi-Tenant Support

Manage agent fleets across multiple teams, environments, and clients from a single AICM instance.

Agent Integrity & Compromise Monitor