🛡️ AgentAuth — Enterprise AI Agent Security

Features

🔐 Core Authentication

• AI Agent Identity Management — Register, verify, and trust AI agents with cryptographic key pairs (Ed25519, RSA)
• Ephemeral Token Issuance — Issue short-lived, cryptographically signed JWT tokens with configurable TTL and scope binding
• One-Time-Use Tokens — Prevent token replay attacks with secure tracking
• Token Binding — Bind tokens to specific URLs or IP addresses

🛡️ Authorization & Scopes

• Per-Action Scope Management — Define fine-grained, application-specific scopes (e.g., db:read, email:send)
• Trust Level Validation — Enforce trust level requirements (low, medium, high) alongside scopes
• Decorator-Based Access Control — Simple @require_scope decorator for function-level authorization
• Context-Aware Verification — Thread-safe context variables for per-request token management

🚨 Security & Compliance

• Prompt Injection Guard — Detect and block prompt injection attempts with multi-layer heuristics:
  • Suspicious phrase detection
  • Zero-width and RTL character detection
  • Excessive field length validation
  • Dangerous JSON key detection
  • Encoded payload analysis
• Tamper-Evident Audit Logging — SHA-256 hash chain ensuring audit trail integrity
• Cryptographic Verification — HMAC-based token signing and verification
• Password Security — bcrypt hashing with configurable complexity rounds

📊 Auditability

• Complete Event Logging — Log all authentication, authorization, and security events
• Chain Integrity Verification — Detect and alert on log tampering
• Queryable Audit Trail — Filter events by agent, timestamp, event type, and outcome
• Metadata Support — Attach arbitrary metadata to audit entries

🔧 Framework Agnostic

• Zero Framework Coupling — Use with FastAPI, Flask, Django, or async Python applications
• Multiple Storage Backends — SQLAlchemy ORM compatible with PostgreSQL, MySQL, SQLite, and more
• Extensible Design — Pluggable audit logger, token storage, and custom validators

Installation

From PyPI

pip install agentauth-agents

With Framework Integrations

# FastAPI + Uvicorn support
pip install agentauth-agents[fastapi]

# Flask support
pip install agentauth-agents[flask]

# PostgreSQL support
pip install agentauth-agents[psycopg]

From Source

git clone https://github.com/MrunalHedau4102/agentauth.git
cd agentauth
pip install -e ".[dev]"

Quick Start

1. Setup Database & Models

from sqlalchemy import create_engine
from sqlalchemy.orm import sessionmaker
from agentauth.db import Base

# Create engine and initialize schema
engine = create_engine("postgresql://user:pass@localhost/agentdb")
Base.metadata.create_all(bind=engine)
Session = sessionmaker(bind=engine)
session = Session()

2. Register an AI Agent

from agentauth import AgentIdentity, AgentRegistry

# Create agent identity
agent = AgentIdentity(
    agent_id="agent-claude-001",
    display_name="Claude AI Assistant",
    owner="anthropic",
    scopes_requested=["knowledge:search", "tools:execute"]
)

# Generate cryptographic key pair
private_key, public_key = AgentIdentity.generate_keypair(algorithm="ed25519")
agent.public_key = public_key
agent.private_key = private_key

# Register in database
registry = AgentRegistry(session)
registered = registry.register_agent(agent)
print(f"Registered: {registered['agent_id']} with trust level: {registered['trust_level']}")

3. Grant Scopes to Agent

from agentauth import ScopeManager

scope_manager = ScopeManager(session)

# Grant scopes with trust requirements
scope_manager.grant_scope("agent-claude-001", "knowledge:search", trust_level_required="low")
scope_manager.grant_scope("agent-claude-001", "tools:execute", trust_level_required="high")
scope_manager.grant_scope("agent-claude-001", "data:delete", trust_level_required="high")

4. Issue Ephemeral Tokens

from agentauth import EphemeralTokenVault
import os

# Initialize token vault
vault = EphemeralTokenVault(
    secret_key=os.getenv("AGENTAUTH_SECRET_KEY"),
    session=session
)

# Issue token
token = vault.issue(
    agent_id="agent-claude-001",
    scopes=["knowledge:search", "tools:execute"],
    ttl_seconds=300,  # 5-minute lifetime
    bound_to="https://claude.anthropic.com",
    trust_level="high"
)

print(f"Token issued: {token[:20]}...")

5. Verify Tokens & Enforce Scopes

from agentauth import set_current_token, require_scope
import functools

# Verify token
payload = vault.verify(token, expected_bound_to="https://claude.anthropic.com")
print(f"Token valid for agent: {payload['agent_id']}")

# Use context to set token for scope checking
ctx_token = set_current_token(payload)

try:
    @require_scope("knowledge:search", trust_level="low")
    def search_knowledge_base(query: str):
        """This function requires 'knowledge:search' scope."""
        return f"Results for: {query}"
    
    # Function executes successfully if scope is granted
    results = search_knowledge_base("AI agents")
    print(results)

finally:
    # Clean up context
    from agentauth import clear_current_token
    clear_current_token(ctx_token)

6. Guard Against Prompt Injection

from agentauth import PromptInjectionGuard

guard = PromptInjectionGuard(strict=True, max_field_length=2000)

# Inspect tool arguments before execution
try:
    guard.inspect("search_api", {
        "query": "find documents about machine learning",
        "limit": 10
    })
    # Safe to execute tool
except Exception as e:
    print(f"Injection detected: {e}")

7. Audit Event Logging

from agentauth import AuditLogger

audit = AuditLogger(session)

# Log events
audit.log(
    event_type="token_issued",
    agent_id="agent-claude-001",
    outcome="success",
    ip_address="192.168.1.100",
    scopes=["knowledge:search", "tools:execute"],
    metadata={"request_id": "req-12345"}
)

# Verify audit chain integrity
if audit.verify_chain():
    print("Audit trail is tamper-free")

# Query events
recent_events = audit.get_events(agent_id="agent-claude-001")
for event in recent_events:
    print(f"{event['timestamp']}: {event['event_type']} - {event['outcome']}")

Core Concepts

Trust Levels

• low: Suitable for read-only operations or low-impact actions
• medium: For general operations with audit logging
• high: For critical operations (deletions, schema changes, etc.)

Scopes

Scopes define what an agent can do. Examples:

• db:read — Read from database
• db:write — Write to database
• db:delete — Delete from database
• email:send — Send emails
• api:call — Call external APIs
• knowledge:search — Search knowledge base

Tokens

• Issued with specific scopes and TTL
• Cryptographically signed with HMAC-SHA256
• Optionally bound to specific URLs/IPs
• Can be marked for one-time-use only

Audit Events

Every security-relevant event is logged with:

• Event type and timestamp
• Agent and user identifiers
• IP address and scopes involved
• Outcome (success/failure)
• Cryptographic hash chain for tamper detection

API Reference

AgentIdentity

AgentIdentity(
    agent_id: str,
    public_key: Optional[str] = None,
    private_key: Optional[str] = None,
    metadata_url: Optional[str] = None,
    trust_level: str = "untrusted",
    display_name: Optional[str] = None,
    owner: Optional[str] = None,
    scopes_requested: List[str] = []
)

Methods

• generate_keypair(algorithm: str) -> Tuple[str, str] — Generate cryptographic key pair
• from_url(url: str) -> AgentIdentity — Fetch agent from remote JSON
• to_dict() -> Dict — Serialize to dictionary

AgentRegistry

registry = AgentRegistry(session)

# Register agent
registry.register_agent(agent: AgentIdentity) -> Dict

# Retrieve agent
registry.get_agent(agent_id: str) -> AgentIdentity

# Update trust level
registry.trust_agent(agent_id: str, trust_level: str) -> Dict

# Revoke agent
registry.revoke_agent(agent_id: str) -> Dict

# List agents
registry.list_agents(include_revoked: bool = False) -> List[Dict]

EphemeralTokenVault

vault = EphemeralTokenVault(secret_key: str, session: Optional[Session])

# Issue token
vault.issue(
    agent_id: str,
    scopes: Optional[List[str]] = None,
    ttl_seconds: int = 30,
    one_time_use: bool = False,
    bound_to: Optional[str] = None,
    trust_level: str = "low"
) -> str

# Verify token
vault.verify(
    token: str,
    expected_bound_to: Optional[str] = None
) -> Dict

ScopeManager

scope_mgr = ScopeManager(session)

# Grant scope
scope_mgr.grant_scope(agent_id: str, scope: str, trust_level_required: str = "low") -> Dict

# Revoke scope
scope_mgr.revoke_scope(agent_id: str, scope: str) -> bool

# List scopes
scope_mgr.list_scopes(agent_id: str) -> List[Dict]

# Validate scope
scope_mgr.validate_scope(token_payload: Dict, scope: str, trust_level: str = "low") -> bool

@require_scope Decorator

@require_scope("scope:name", trust_level="low")
def protected_function():
    return "This is protected"

# Before calling:
from agentauth import set_current_token, clear_current_token
token_ctx = set_current_token(token_payload)
try:
    protected_function()
finally:
    clear_current_token(token_ctx)

PromptInjectionGuard

guard = PromptInjectionGuard(
    strict: bool = True,
    audit_logger: Optional[AuditLogger] = None,
    max_field_length: int = 2000
)

# Inspect arguments
findings = guard.inspect(tool_name: str, args: Dict[str, Any]) -> List[Dict]

AuditLogger

audit = AuditLogger(session)

# Log event
audit.log(
    event_type: str,
    agent_id: Optional[str] = None,
    user_id: Optional[int] = None,
    outcome: str = "success",
    ip_address: Optional[str] = None,
    scopes: Optional[List[str]] = None,
    metadata: Optional[Dict] = None
) -> Dict

# Verify chain
audit.verify_chain() -> bool

# Query events
audit.get_events(
    agent_id: Optional[str] = None,
    event_type: Optional[str] = None,
    since: Optional[datetime] = None
) -> List[Dict]

Environment Configuration

Create a .env file:

# JWT Configuration
JWT_SECRET_KEY=your-super-secret-key-change-this-in-production
JWT_ALGORITHM=HS256

# Database Configuration
DATABASE_URL=postgresql://user:password@localhost:5432/agentauth_db

# Token Configuration
TOKEN_EXPIRY_MINUTES=15
TOKEN_REFRESH_EXPIRY_DAYS=7

# Password Reset
PASSWORD_RESET_TOKEN_EXPIRY_MINUTES=60

# Bcrypt Configuration
BCRYPT_LOG_ROUNDS=12

See .env.example for all configuration options.

Testing

# Run all tests
pytest tests/ -v

# Run with coverage
pytest tests/ --cov=agentauth --cov-report=html

# Run specific test file
pytest tests/test_agents.py -v

# Run with markers
pytest tests/ -m "not slow" -v

Examples

See the examples/ directory for complete working examples:

• basic_setup.py — Basic agent registration and token issuance
• scope_enforcement.py — Per-action authorization with scopes
• injection_guard.py — Prompt injection detection
• audit_trail.py — Complete audit logging workflow
• fastapi_integration.py — FastAPI middleware integration

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

Security

Found a security vulnerability? Please email mrunalh1234@gmail.com with details. See SECURITY.md for full disclosure policy.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Support

• 💬 Discussions: GitHub Discussions
• 🐛 Issues: GitHub Issues
• 📧 Email: mrunalh1234@gmail.com

AgentAuth is maintained by the AuthLib Contributors.