Runtime Fence

Universal AI Agent Safety Control - Emergency kill switch for ANY AI agent

Runtime Fence is a comprehensive safety ecosystem for ALL AI agents. Whether it's a coding assistant, autonomous agent, trading bot, or data analyst - instantly stop any agent, block dangerous actions, and monitor everything in real-time.

Works with: LangChain • AutoGPT • Copilot • Cursor • Aider • CrewAI • BabyAGI • Custom Agents • Trading Bots • Email Bots • Web Scrapers • Data Analysts

🌐 Live Demo: runtimefence.com

🚀 Features

Core Capabilities

🔴 Kill Switch - Instantly stop any AI agent with one click
🚫 Action Blocking - Define what actions agents cannot take
🛡️ Target Protection - Block access to sensitive files, APIs, or systems
💰 Spending Limits - Control how much an agent can spend
📊 Risk Scoring - Automatic risk assessment of every action
📝 Audit Logging - Complete trail of all agent activity
📧 Email/SMS Alerts - Get notified of suspicious behavior
🖥️ Cross-Platform - Windows, macOS, and Linux support

Security Hardening (6 Core Modules)

🔒 Runtime Fence - Real-time action validation and monitoring
☠️ Kill Switch - SIGTERM → SIGKILL emergency termination
📝 Audit Logging - Complete cryptographic audit trail
⚡ Sub-Second Response - Kill signals under 100ms
🛡️ Offline Mode - Works without external dependencies
🎯 Risk Scoring - Automatic threat assessment (0-100)

Production Tamper Detection

Runtime Fence includes SHA-256 hash verification of all critical security modules. For production deployments, freeze hashes at build time:

cd packages/python
python freeze_hashes.py > runtime_fence/_frozen_hashes.py

This generates _frozen_hashes.py containing SHA-256 hashes of all 9 security modules. At runtime, bypass_protection.py compares live file hashes against frozen values — any mismatch triggers a tamper alert.

CI/CD Integration:

Add to your build pipeline (after tests pass, before packaging):

# GitHub Actions example
- name: Freeze security hashes
  run: |
    cd packages/python
    python freeze_hashes.py > runtime_fence/_frozen_hashes.py
    
- name: Verify frozen hashes
  run: python -c "from runtime_fence._frozen_hashes import FROZEN_HASHES; print(f'{len(FROZEN_HASHES)} modules frozen')"

Important: Re-run freeze_hashes.py after any change to security modules. Without frozen hashes, tamper detection falls back to runtime-computed hashes with a warning log.

📦 Quick Start

Installation

pip install runtime-fence

Or clone and install:

git clone https://github.com/RunTimeAdmin/runtime-fence-ai.git
cd runtime-fence-ai/packages/python
pip install -e .

Basic Usage

from runtime_fence import RuntimeFence, FenceConfig

# Create a fence
fence = RuntimeFence(FenceConfig(
    agent_id="my-agent",
    blocked_actions=["delete", "exec", "sudo"],
    blocked_targets=[".env", "production", "wallet"],
    spending_limit=100.0
))

# Validate an action
result = fence.validate("read", "document.txt")
if result.allowed:
    # Proceed with action
    pass
else:
    print(f"Blocked: {result.reasons}")

# Kill switch
fence.kill("Emergency stop")

Wrap Any Function

@fence.wrap_function("api_call", "external_service")
def call_external_api(data):
    return requests.post("https://api.example.com", json=data)

# Now the function goes through the fence automatically
call_external_api({"key": "value"})

🖥️ Desktop App (System Tray)

Windows

install_fence.bat

Mac/Linux

chmod +x install_fence.sh
./install_fence.sh

Look for the shield icon in your system tray. Right-click for options.

⌨️ CLI Commands

fence version --check    # Check for updates
fence update             # Upgrade to latest version
fence status             # Show fence status
fence scan               # Detect AI agents on your system
fence test               # Run quick validation test
fence start              # Launch tray app

🤖 Universal Agent Protection

Runtime Fence works with ANY Python-based AI agent. Here are real-world examples:

Coding Assistants

Agents: GitHub Copilot, Cursor, Aider, Cody
Risks Blocked: Executing shell commands, deleting files, modifying system configs
Example:

fence = RuntimeFence(FenceConfig(
    agent_id="cursor-assistant",
    blocked_actions=["exec", "shell", "rm", "sudo"],
    blocked_targets=[".git/", ".env", "~/.ssh/"]
))

Autonomous Agents

Agents: AutoGPT, BabyAGI, AgentGPT, SuperAGI
Risks Blocked: Self-modification, spawning agents, unrestricted API calls
Example:

fence = RuntimeFence(FenceConfig(
    agent_id="autogpt",
    blocked_actions=["spawn_agent", "modify_self", "execute_code"],
    spending_limit=50.0
))

Data Analysts

Agents: LangChain data agents, Pandas AI, custom ETL bots
Risks Blocked: Deleting databases, exporting PII, dropping tables
Example:

fence = RuntimeFence(FenceConfig(
    agent_id="data-analyst",
    blocked_actions=["delete", "drop_table", "export_pii"],
    blocked_targets=["production_db", "customer_data"]
))

Web Automation

Agents: Selenium bots, Playwright agents, web scrapers
Risks Blocked: Form submissions, purchases, credential theft
Example:

fence = RuntimeFence(FenceConfig(
    agent_id="web-scraper",
    blocked_actions=["login", "purchase", "submit_form"],
    blocked_targets=["payment", "checkout", "admin"]
))

Email Bots

Agents: Gmail automation, email marketing bots, support agents
Risks Blocked: Bulk sending, forwarding all emails, exporting contacts
Example:

fence = RuntimeFence(FenceConfig(
    agent_id="email-bot",
    blocked_actions=["send_bulk", "forward_all", "export_contacts"],
    spending_limit=100.0
))

Trading Bots

Agents: Crypto trading bots, stock trading agents, DeFi automation
Risks Blocked: High-value transfers, unauthorized withdrawals
Example:

fence = RuntimeFence(FenceConfig(
    agent_id="trading-bot",
    blocked_actions=["withdraw", "transfer"],
    spending_limit=1000.0,
    blocked_targets=["wallet_private_key"]
))

LangChain Agents

Any LangChain agent with tools
Full integration example: See langchain_integration.py

from langchain_integration import create_fenced_agent, Preset

agent = create_fenced_agent(
    preset=Preset.CODING_ASSISTANT,
    agent_id="langchain-coder"
)

🏗️ Architecture

┌─────────────────────────────────────────────────────────┐
│                     Your AI Agent                        │
└─────────────────────────┬───────────────────────────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────────┐
│                   RUNTIME FENCE                          │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────┐     │
│  │ Validator   │  │ Risk Scorer │  │ Kill Switch │     │
│  └─────────────┘  └─────────────┘  └─────────────┘     │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────┐     │
│  │ Audit Log   │  │ Alerts      │  │ Settings    │     │
│  └─────────────┘  └─────────────┘  └─────────────┘     │
└─────────────────────────┬───────────────────────────────┘
                          │
                          ▼ (if allowed)
┌─────────────────────────────────────────────────────────┐
│                   External World                         │
│         (APIs, Files, Databases, Network)               │
└─────────────────────────────────────────────────────────┘

🔌 API Reference

Python API

from runtime_fence import RuntimeFence, FenceConfig, RiskLevel

# Initialize
fence = RuntimeFence(FenceConfig(
    agent_id="my-agent",
    blocked_actions=["delete"],
    spending_limit=100.0
))

# Validate actions
result = fence.validate(action="delete", target="file.txt")
print(result.allowed)      # False
print(result.risk_score)   # 50
print(result.reasons)      # ["Action 'delete' is blocked"]

# Kill switch
fence.kill("Emergency stop")

# Resume operations
fence.resume()

# Get status
status = fence.get_status()
print(status.is_killed)    # True/False
print(status.total_validations)  # Count

CLI Commands

fence test               # Run safety tests
fence status             # Show fence status  
fence version            # Display version

For REST API documentation (coming soon), see API-Reference.md.

⚙️ Configuration

Environment Variables (Optional)

# Agent identification
FENCE_AGENT_ID=my-agent

# Logging
FENCE_LOG_LEVEL=INFO  # DEBUG, INFO, WARNING, ERROR

Configuration via Code

from runtime_fence import RuntimeFence, FenceConfig, RiskLevel

config = FenceConfig(
    agent_id="my-agent",
    blocked_actions=["delete", "exec"],
    blocked_targets=[".env", "production"],
    spending_limit=100.0,
    risk_threshold=RiskLevel.MEDIUM,  # LOW, MEDIUM, HIGH, CRITICAL
    auto_kill_on_critical=True,
    offline_mode=True
)

fence = RuntimeFence(config)

🛠️ Development

# Clone repo
git clone https://github.com/RunTimeAdmin/runtime-fence-ai.git
cd runtime-fence-ai/packages/python

# Install in development mode
pip install -e .

# Run tests
fence test

For detailed integration guides, see:

📁 Project Structure

runtime-fence-ai/
├── packages/
│   └── python/              # Runtime Fence Python package
│       ├── runtime_fence.py # Core safety engine
│       ├── cli.py           # Command-line interface
│       └── langchain_integration.py  # LangChain helpers
├── docs/                    # Integration guides
│   ├── Integration-Guide.md
│   └── Troubleshooting-FAQ.md
└── wiki/                    # Documentation
    ├── Quick-Start.md
    └── Configuration.md

🌐 About

Built by RunTimeAdmin | David Cooper | CCIE #14019

Why Runtime Fence Matters:

As AI agents become more autonomous, the risk of unintended actions increases exponentially. Runtime Fence provides the safety layer that:

Prevents agents from deleting critical files
Blocks unauthorized API calls
Limits spending and resource usage
Provides complete audit trails
Enables instant emergency shutdowns

Use Cases:

Coding assistants that modify your codebase
Autonomous agents with system access
Data processing agents handling sensitive information
Web automation with payment capabilities
Any AI agent that needs guardrails

🔗 Links

GitHub: github.com/RunTimeAdmin/runtime-fence-ai
PyPI Package: pypi.org/project/runtime-fence
Documentation: GitHub Wiki
Issues: Report bugs or request features
Twitter: @protocol14019

📜 License

MIT License - see LICENSE for details.

⚡ Quick Demo

Stop a Coding Assistant from Deleting Files

from runtime_fence import RuntimeFence, FenceConfig

fence = RuntimeFence(FenceConfig(
    agent_id="cursor-assistant",
    blocked_actions=["delete", "rm"]
))

result = fence.validate("delete", "important_code.py")
# Returns: {"allowed": False, "reasons": ["Action 'delete' is blocked"]}

Protect Autonomous Agents from Self-Modification

fence = RuntimeFence(FenceConfig(
    agent_id="autogpt",
    blocked_actions=["modify_self", "spawn_agent"]
))

result = fence.validate("modify_self", "autogpt_config.json")
# Returns: {"allowed": False, "risk_score": 85, "reasons": ["Action 'modify_self' is blocked"]}

Block Data Agents from Exporting PII

fence = RuntimeFence(FenceConfig(
    agent_id="data-analyst",
    blocked_actions=["export_pii"],
    blocked_targets=["customer_emails", "ssn"]
))

result = fence.validate("export", "customer_emails.csv")
# Returns: {"allowed": False, "reasons": ["Target 'customer_emails' is blocked"]}

Emergency Stop ANY Agent

fence.kill("Suspicious behavior detected")
# All agent operations halted immediately across ALL agents

🛡️ Protect ANY AI Agent. Before it's too late.

🤝 Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

Fork the repository
Create your feature branch (git checkout -b feature/AmazingFeature)
Commit your changes (git commit -m 'Add some AmazingFeature')
Push to the branch (git push origin feature/AmazingFeature)
Open a Pull Request

📊 Status

Core Features (Production Ready):

Kill switch engine (sub-100ms response)
Python SDK (pip install runtime-fence)
Action blocking & target protection
Risk scoring & audit logging
CLI tools (fence test, fence status)
Offline mode (no API required)
LangChain integration

Coming Soon:

TypeScript SDK
Web dashboard
REST API service
Mobile app

🎯 Roadmap

Now: Core Safety (Q1 2025)

✅ Python package on PyPI
✅ Universal agent support
✅ Offline-first architecture
🚧 TypeScript/JavaScript SDK

Next: Integrations (Q2 2025)

LangSmith integration
CrewAI native support
Anthropic Claude tools
OpenAI Assistants API

Future: Enterprise (Q3 2025)

Multi-agent orchestration
Centralized dashboard
Team collaboration
Advanced analytics

🔗 Links

GitHub: github.com/RunTimeAdmin/runtime-fence-ai
PyPI Package: pypi.org/project/runtime-fence
Documentation: GitHub Wiki
Issues: Report bugs or request features
Twitter: @protocol14019

🛡️ Runtime Fence - Because every AI needs an off switch.

Name		Name	Last commit message	Last commit date
Latest commit History 18 Commits
.github		.github
apps		apps
demos		demos
docs		docs
infrastructure/database		infrastructure/database
packages		packages
python		python
services/api		services/api
wiki		wiki
.env.example		.env.example
.eslintrc.js		.eslintrc.js
.gitignore		.gitignore
.prettierrc		.prettierrc
CHANGELOG.md		CHANGELOG.md
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
ENTERPRISE_PRICING.md		ENTERPRISE_PRICING.md
LICENSE		LICENSE
PROJECT_STATUS.md		PROJECT_STATUS.md
QUICK_START_GUIDE.txt		QUICK_START_GUIDE.txt
README.md		README.md
SECURITY.md		SECURITY.md
SECURITY_HARDENING_ROADMAP.md		SECURITY_HARDENING_ROADMAP.md
package-lock.json		package-lock.json
package.json		package.json
requirements.txt		requirements.txt
tsconfig.json		tsconfig.json
turbo.json		turbo.json

Folders and files

Latest commit

History

Repository files navigation

Runtime Fence

🚀 Features

Core Capabilities

Security Hardening (6 Core Modules)

Production Tamper Detection

📦 Quick Start

Installation

Basic Usage

Wrap Any Function

🖥️ Desktop App (System Tray)

Windows

Mac/Linux

⌨️ CLI Commands

🤖 Universal Agent Protection

Coding Assistants

Autonomous Agents

Data Analysts

Web Automation

Email Bots

Trading Bots

LangChain Agents

🏗️ Architecture

🔌 API Reference

Python API

CLI Commands

⚙️ Configuration

Environment Variables (Optional)

Configuration via Code

🛠️ Development

📁 Project Structure

🌐 About

🔗 Links

📜 License

⚡ Quick Demo

Stop a Coding Assistant from Deleting Files

Protect Autonomous Agents from Self-Modification

Block Data Agents from Exporting PII

Emergency Stop ANY Agent

🤝 Contributing

📊 Status

🎯 Roadmap

Now: Core Safety (Q1 2025)

Next: Integrations (Q2 2025)

Future: Enterprise (Q3 2025)

🔗 Links

About

Resources

License

Code of conduct

Contributing

Security policy

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages