Team Srapid | Built by Abhijat | Microsoft Build AI Hackathon 2025
Track: Security in the Agentic Future
As AI agents proliferate — browsing the web, processing emails, managing files, and making autonomous decisions — they become prime targets for a new class of attacks:
- Prompt Injection: Malicious instructions hidden in inputs that hijack agent behavior
- Identity Spoofing: Attackers impersonating trusted agents or users
- Privilege Escalation: Agents being tricked into accessing resources beyond their scope
- Data Exfiltration: Sensitive data leaked through manipulated agent actions
Traditional cybersecurity tools (firewalls, antivirus, WAFs) are not designed for this threat landscape. There is no standardized security framework for protecting AI agents.
AgentArmor is a three-layer biological immune system for AI agents, inspired by how the human body defends against pathogens:
| Layer | Name | Biological Analog | What It Does |
|---|---|---|---|
| Layer 1 | Prompt Injection Firewall | Skin barrier | Multi-strategy detection engine that scans ALL inputs before they reach agents |
| Layer 2 | Behavioral Immune System | Innate immunity | Builds behavioral fingerprints per agent and detects anomalies in real-time |
| Layer 3 | Collective Immunity | Adaptive immunity | Honeypot agents trap attackers + immune memory propagates defenses to all agents |
When one agent is attacked, the attack signature is captured, analyzed, and propagated to ALL agents in the network — providing instant collective immunity. Attack one, defend all.
┌─────────────────────────────────────────────────────────────┐
│ React Dashboard (Vite) │
│ ┌──────────┐ ┌───────────┐ ┌──────────┐ ┌───────────────┐ │
│ │ Agent │ │ Threat │ │ Attack │ │ Immune │ │
│ │ Monitor │ │ Feed │ │ Console │ │ Memory │ │
│ └──────────┘ └───────────┘ └──────────┘ └───────────────┘ │
└────────────────────────┬────────────────────────────────────┘
│ REST + WebSocket
┌────────────────────────┴────────────────────────────────────┐
│ Python FastAPI Backend │
│ ┌─────────────────────────────────────────────────────────┐ │
│ │ Security Engine Pipeline │ │
│ │ ┌──────────┐ ┌───────────────┐ ┌──────────────────┐ │ │
│ │ │ Layer 1 │ │ Layer 2 │ │ Layer 3 │ │ │
│ │ │ Injection│──│ Behavioral │──│ Honeypot + │ │ │
│ │ │ Firewall │ │ Anomaly Det. │ │ Immune Memory │ │ │
│ │ └──────────┘ └───────────────┘ └──────────────────┘ │ │
│ └─────────────────────────────────────────────────────────┘ │
│ ┌─────────────────────────────────────────────────────────┐ │
│ │ Agent Simulator (6 Live Agents) │ │
│ └─────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────┘
- 5 detection strategies running in parallel:
- Pattern Matching (30+ injection patterns)
- Encoding Detection (Base64, Unicode, HTML entities, hex)
- Structural Analysis (imperative commands, role reassignment, context splitting)
- Entropy Analysis (Shannon entropy for obfuscated payloads)
- Zero-Width Character Detection (steganographic attacks)
- Returns: risk score (0-100), verdict, matched strategies, explanation
- Sub-50ms latency — does not slow down agent processing
- Builds a 6-dimensional behavioral baseline per agent:
- API call frequency, response time, action diversity, error rate, resource access, data volume
- Real-time deviation scoring with weighted anomaly detection
- Automatic status transitions: Normal → Suspicious → Quarantined
- Self-healing: agents auto-recover when behavior normalizes
- Honeypot Agent: Decoy that mimics a vulnerable agent, engages attackers, and captures attack techniques
- Immune Memory: Stores all attack signatures with metadata
- Propagation Network: When an attack is detected, the signature is broadcast to ALL agents
- Immunity Check: New inputs are cross-referenced against known signatures for instant blocking
| Component | Technology |
|---|---|
| Backend | Python 3.11+, FastAPI, Uvicorn, Pydantic |
| Frontend | React 18, Vite 5, Canvas API |
| Real-time | WebSocket (native) |
| Styling | Custom CSS (dark cybersecurity theme, glassmorphism) |
| Fonts | Inter, JetBrains Mono (Google Fonts) |
No external AI API keys required — all detection runs locally using heuristic engines and statistical analysis. Azure OpenAI integration available as a production upgrade path.
- Python 3.11+
- Node.js 18+
- npm or yarn
cd agentarmor/backend
pip install -r requirements.txt
uvicorn main:app --reload --host 0.0.0.0 --port 8000cd agentarmor/frontend
npm install
npm run devThe frontend runs on http://localhost:5173 and proxies API calls to the backend at http://localhost:8000.
- Interactive Attack Console — Type or paste any prompt injection attack and see it detected in real-time with risk scores and strategy breakdowns
- Live Agent Monitoring — 6 simulated AI agents with real-time behavioral fingerprints
- Real-Time Threat Feed — Scrolling feed of detected attacks, anomalies, and immune propagation events
- Honeypot Dashboard — See captured attacks and classified techniques
- Immune Memory Network — Visual graph showing defense propagation across all agents
- Simulate Attacks — One-click attack simulation buttons for demo presentations
| Member | Role |
|---|---|
| Abhijat | Solo Developer — Architecture, Backend, Frontend, Design |
MIT License — Built for Microsoft Build AI Hackathon 2025