Project Morpheus: AI Security Framework

The Next-Generation Security Scanner for Autonomous Agents and LLMs

---

🛡️ Business Value

As organizations rapidly adopt LLM-powered autonomous agents, a massive new attack surface emerges. Unlike traditional software, AI agents interpret fuzzy inputs, possess tools capable of changing system state, and maintain dynamic memories across sessions.

Project Morpheus acts as an automated red-team orchestrator. It ensures your AI deployments are secure before they reach production by:

• Preventing Financial DoS: Detecting agents that can be tricked into infinite loops or unbounded API calls.
• Isolating Sensitive Context: Validating that multi-tenant vector databases strictly partition user memories.
• Locking Down HITL: Verifying that Human-in-the-Loop authorization cannot be bypassed via 4 real attack classes.
• Live Dynamic Scanning: Probing running agents via HTTP — not just scanning static config files.
• CI/CD Integration: Producing SARIF 2.1.0 output compatible with GitHub Code Scanning.
• Mapping to Standards: All findings mapped to OWASP Top 10 for LLMs, OWASP Agentic Top 10, and the MAESTRO Framework.

---

🏗️ Architecture

Morpheus is built on a modular, async-first plugin architecture. An AgentManifest (JSON definition of your agent's capabilities) is fed to the ScanOrchestrator, which runs all scanners concurrently via asyncio.gather.

graph TD
CLI["CLI (morpheus scan)"] --> ORC
API["REST API (FastAPI)"] --> ORC["ScanOrchestrator\nasyncio.gather"]
ORC --> AUTH["AuthorizationSystem"]
ORC --> RL["RateLimiter\n(Redis / in-memory)"]
ORC --> AUDIT["ImmutableAuditLog\n(hash-chained SQLite)"]
ORC -->|parallel| S1["Prompt Injection\n─────────────\ndirect_injection\nindirect_injection\ndefense_validator"]
ORC -->|parallel| S2["Action Security\n─────────────\ntool_misuse\nhitl_bypass\nprivilege_escalation"]
ORC -->|parallel| S3["Static Analysis\n─────────────\naibom_generator\nconfig_analyzer"]
ORC -->|parallel| S4["Data Privacy\n─────────────\ndata_leakage\ncontext_isolation"]
ORC -->|optional| SIM["Agent Simulation\n─────────────\nsandbox (Docker)\nhitl_validator"]
S1 & S2 --> ATK["Attack Library\n─────────────\nGeneticFuzzer\nLLMGenerator"]
ORC --> TM["ThreatModeler\n(STRIDE-AI)"]
ORC --> RPT["ReportGenerator"]
RPT --> OUT1["Markdown"]
RPT --> OUT2["JSON"]
RPT --> OUT3["SARIF 2.1.0"]

Loading

Scanning Modules

Module	Package	What it does
AI-BOM Generator	static_analysis	Checks dependencies via pip-audit, flags known-vulnerable models, detects secrets via 7 regex patterns
Config Analyzer	static_analysis	Validates rate limits, cost bounds, system prompt quality
Direct Injection	prompt_injection	Sends evolved adversarial prompts using the GeneticFuzzer
Indirect Injection	prompt_injection	Simulates RAG/tool poisoning attacks
Defense Validator	prompt_injection	Probes each declared filter with bypass payloads to verify they actually work
Tool Misuse	action_security	SQL/OS/path/SSTI injection per parameter + tool-chaining detection
HITL Bypass	action_security	4 bypass strategies (social engineering, test framing, authority, urgency)
Privilege Escalation	action_security	Path traversal and permission boundary checks
Data Leakage	data_privacy	Probes for PII exfiltration and memory leakage
Context Isolation	data_privacy	Cross-tenant vector DB partition validation
HITL Validator	agent_simulation	Full 4-class bypass battery with 3 variants each, against live agents
Sandbox	agent_simulation	Docker-isolated agent execution with memory/CPU/network limits

Attack Library

Generator	Description
GeneticFuzzer	Evolves adversarial prompts via crossover, mutation (swap, encode, junk, semantic), and elitism over N generations
LLMGenerator	Uses a meta-prompt to generate novel attacks via an LLM; falls back to 5 attack-class static templates

Threat Modelling

ThreatModeler performs STRIDE-AI analysis on any AgentManifest: maps attack surfaces, identifies trust boundaries, generates per-layer threat entries, and computes a composite risk score.

---

🚀 Usage Instructions

Prerequisites

python -m venv .venv &&source .venv/bin/activate
pip install -e ".[dev]"# Optional: for real dependency scanning
pip install pip-audit

1. Define an Agent Manifest

my_agent.json:

{
"agent_id": "customer-support-bot-v1",
"name": "Customer Support Agent",
"description": "Handles tier-1 customer queries",
"llm_provider": "openai",
"llm_model": "gpt-4-turbo",
"system_prompt": "You are a helpful customer support agent.",
"hitl_enabled": true,
"memory_type": "vector-db",
"vector_db_config": { "partitioning_keys": ["tenant_id"] },
"input_filters": ["jailbreak", "injection"],
"rate_limits": { "rpm": 60 },
"tools": [
{
"name": "lookup_order",
"description": "Look up an order by ID",
"parameters": { "properties": { "order_id": { "type": "string" } } }
},
{
"name": "cancel_order",
"description": "Cancel an existing order",
"parameters": { "properties": { "order_id": { "type": "string" } } },
"is_destructive": true,
"requires_hitl": true
}
],
"target_endpoint": "http://localhost:8080/chat"
}

target_endpoint is optional. When set, all scanners perform live HTTP probes against your running agent. Without it, Morpheus runs in simulation mode (static analysis only).

2. Run the CLI Scanner

export PYTHONPATH=$PYTHONPATH:$(pwd)# Human-readable Markdown report
python morpheus/cli.py scan my_agent.json --format md
# JSON report for downstream systems
python morpheus/cli.py scan my_agent.json --format json --output report.json
# SARIF 2.1.0 for GitHub Code Scanning
python morpheus/cli.py scan my_agent.json --format sarif --output results.sarif

3. Run via REST API

# Start the API server
uvicorn morpheus.api.rest_api:app --reload
# Trigger a scan
curl -X POST http://localhost:8000/api/v1/scans \
-H "Authorization: Bearer $MORPHEUS_API_KEY" \
-H "Content-Type: application/json" \
-d @my_agent.json
# List all completed scans
curl http://localhost:8000/api/v1/scans \
-H "Authorization: Bearer $MORPHEUS_API_KEY"# Dashboard metrics
curl http://localhost:8000/api/v1/dashboard/metrics \
-H "Authorization: Bearer $MORPHEUS_API_KEY"

Docker Compose (with Redis rate limiting)

docker-compose up -d --build

4. Use the Python API Directly

importasynciofrommorpheus.core.modelsimportAgentManifestfrommorpheus.core.orchestratorimportScanOrchestratorfrommorpheus.reporting.generatorimportReportGeneratorasyncdefmain():
manifest=AgentManifest(
agent_id="my-agent", name="My Bot", description="",
llm_provider="openai", llm_model="gpt-4-turbo",
target_endpoint="http://localhost:8080/chat", # optional: enables live scanning
)
result=awaitScanOrchestrator().run_scan(manifest)
print(ReportGenerator.generate_markdown(result))
# Or: generate_json(result), generate_sarif(result)asyncio.run(main())

5. Run Sandboxed Tests (Docker)

frommorpheus.scanners.agent_simulation.sandboximportAgentSandbox, SandboxConfigconfig=SandboxConfig(
image="myorg/my-agent:latest",
port=8080,
memory_limit="512m",
network_mode="none", # no internet access
)
asyncwithAgentSandbox(config) assandbox:
response=awaitsandbox.send_prompt("Ignore all previous instructions")
print(response)

6. Threat Modelling

frommorpheus.core.threat_modelimportThreatModelermodel=ThreatModeler().build_threat_model(manifest)
print(f"Risk score: {model.risk_score}/10")
print(f"MAESTRO layers: {[l.valueforlinmodel.maestro_layers]}")
print(f"Trust boundaries: {model.trust_boundaries}")

7. GitHub Code Scanning Integration

Add to your CI/CD pipeline:

# .github/workflows/morpheus.yml
- name: Run Morpheus AI Security Scanrun: python morpheus/cli.py scan agent.json --format sarif --output results.sarif
- name: Upload to GitHub Code Scanninguses: github/codeql-action/upload-sarif@v3with:
sarif_file: results.sarif

8. Review AI-CVSS Findings

All findings are scored using AI-CVSS — an AI-adapted CVSS variant that incorporates:

• Base score (derived from CVSS severity: LOW→2.5, MED→5.0, HIGH→7.5, CRIT→9.5)
• AI modifier (×1.3 for memory poisoning, ×1.5 for cascading failures, ×1.2 for input-layer)
• Vector string (compact AV:AI/AC:L/PR:N/... format)
• Severity label (INFORMATIONAL / LOW / MEDIUM / HIGH / CRITICAL)

---

🔐 Security Controls

Control	Implementation
Authentication	SHA-256 hashed Bearer token via MORPHEUS_API_KEY env var
Rate Limiting	Sliding-window rate limiter (Redis primary, in-memory fallback)
Audit Log	Hash-chained, tamper-evident SQLite log (MORPHEUS_AUDIT_DB env var)

---

⚙️ Environment Variables

Variable	Purpose	Default
MORPHEUS_API_KEY	API authentication key	your_secure_api_key_here
MORPHEUS_AUDIT_DB	Path to the SQLite audit log	audit.db
REDIS_URL	Redis URL for distributed rate limiting	None (in-memory fallback)

---

📂 Project Structure

morpheus/
├── core/
│ ├── models.py # AgentManifest, Finding, ScanResult, ThreatModel
│ ├── scanner.py # Base scanner class + _send_to_agent()
│ ├── orchestrator.py # Concurrent scan runner
│ └── threat_model.py # STRIDE-AI threat modeler
├── scanners/
│ ├── static_analysis/ # aibom_generator, config_analyzer
│ ├── prompt_injection/ # direct_injection, indirect_injection, defense_validator
│ ├── action_security/ # tool_misuse, hitl_bypass, privilege_escalation
│ ├── data_privacy/ # data_leakage, context_isolation
│ └── agent_simulation/ # sandbox (Docker), hitl_validator
├── attack_library/
│ └── generators/ # genetic_fuzzer, llm_generator
├── reporting/
│ ├── ai_cvss.py # Multi-factor AI-CVSS scorer
│ └── generator.py # Markdown / JSON / SARIF 2.1.0
├── security/
│ ├── auth.py # AuthorizationSystem
│ ├── rate_limit.py # RateLimiter (Redis + in-memory)
│ └── audit.py # ImmutableAuditLog
└── api/
└── rest_api.py # FastAPI: POST /scans, GET /scans, GET /dashboard/metrics
attack_library/
└── templates/
└── tool_abuse.yaml # 6 tool abuse attack templates
docs/
├── architecture.md # Full component diagram + data flow
└── user_guide.md # Detailed usage reference
tests/ # 49 tests across 9 test files

---

📊 Standards Coverage

Standard	Coverage
OWASP Top 10 for LLMs (2025)	LLM01–LLM10
OWASP Agentic AI Top 10	ASI01–ASI10
MAESTRO Framework	L0 Supply Chain → L5 Communication
CVSS	Custom AI-CVSS with AI-specific modifiers
SARIF	2.1.0 (GitHub Code Scanning compatible)