Security First

How We Protect Your Data

Your code is your business. Your data is your asset. Here is exactly how we keep it safe.

🔒

Your Data Stays Yours

We never train models on your data. Your code, conversations, and business information are never shared, sold, or used for AI training. Period.

🏗️

Isolated Infrastructure

Every managed agent runs on a dedicated VPS. No shared databases, no co-tenancy, no cross-customer data leaks. Your agent is your agent.

🔓

No Vendor Lock-In

All files are standard markdown (CLAUDE.md, SKILL.md, memory.md). Export everything at any time. No proprietary formats, no walled gardens.

📖

Open-Source Foundation

Built on open-source tools (Next.js, OpenClaw, Paperclip). You can audit the codebase. No black boxes.

Dual-Gate Skill Review Process

Every third-party skill, tool, and MCP server passes a two-gate review before we include it in our ecosystem. Both gates must pass. No exceptions.

Gate 1: Security

No hardcoded API keys or credentials
No telemetry or data exfiltration
No phone-home behavior or suspicious network calls
No malicious code patterns (eval, exec, subprocess abuse)
License compatible with commercial use

Gate 2: Quality

Actively maintained (recent commits, responsive maintainer)
Documentation exists and is accurate
Relevant to AI Starter Package users
Community traction (stars, forks, issues)
Code quality meets standards

To date: 143 skills reviewed and approved. 3 repositories rejected (telemetry, ToS violation, restrictive license).

Managed Agent Security

How we secure the managed AI agents running on your dedicated VPS.

Secret Management

All API keys stored as environment variables. Never committed to code. Rotated on a regular schedule.

Hook-Based Safety

Automated hooks run before every action. Pre-commit checks prevent pushing to production. Context health monitoring prevents degradation.

Auditor Agent

The Auditor agent reviews work quality before knowledge gets promoted. Human-in-the-loop safety — AI cannot silently modify its own rules.

Knowledge Gate

New learnings go through a nomination → review → promotion pipeline. Not every observation becomes a permanent rule. Quality over quantity.

SSH Key Authentication

Managed agent VPS servers use SSH key authentication only. No password access. Fail2ban blocks brute force attempts.

Firewall Configuration

Only necessary ports are open. UFW configured by default. All unnecessary services disabled.

Automatic Security Updates

Unattended-upgrades enabled on all managed agent servers. Critical patches applied automatically.

Log Monitoring

All agent actions logged to audit trail. Hooks auto-populate logs. Available for review on request.

Runtime Security Layer

Two independent security systems protect your AI agents at runtime — one verifies commands before execution, the other scans for vulnerabilities across the entire system.

Command Firewall

Nexus Gate

Deterministic command firewall — intercepts every shell command before execution. Classifies risk via structural data flow analysis (not AI guessing). Blocks exfiltration, credential leaks, and destructive operations. 195+ known tool patterns, 72 attack signatures.

Zero external dependencies
Self-protection (AI cannot modify Nexus files)
Taint tracking across multi-step attacks
Credential detection (ghp_, sk-proj-, AKIA patterns)
SHA256 behavioral fingerprinting
Hashed audit logs (secrets never written to disk)

Agent Security

Ruflo Security Engine

Automated vulnerability scanning, CVE detection, threat modeling, and input validation. Runs as part of the swarm orchestration layer — every agent action passes through security boundaries.

CVE and vulnerability scanning
Threat modeling for agent architectures
Input validation at all system boundaries
Security-aware agent spawning
Compliance checks (CWE tracking)
Scheduled security scan automation

MCP Server Security

MCP servers connect your AI to external tools. Here is how we manage the risks.

Risk	How We Mitigate
Malicious MCP servers	All MCP servers in our ecosystem are reviewed. We publish the dual-gate review for every server we recommend.
Data leaking to external APIs	MCP servers connect to services you explicitly authorize. No hidden network calls. You control which tools are active.
Prompt injection via tools	Hook-based safety checks run before and after tool use. The Auditor agent flags suspicious patterns.
API key exposure	Keys stored in environment variables, never in CLAUDE.md or settings.json. OS-level secret storage recommended (macOS Keychain, etc.).

What We Will Never Do

✕Train AI models on your data, code, or conversations
✕Share your information with third parties
✕Store your API keys on our servers (you manage your own)
✕Access your managed agent VPS without your permission
✕Lock you into proprietary formats (everything is standard markdown)

Report a Security Issue

Found a vulnerability? Please report it responsibly. We take every report seriously and respond within 24 hours.

security@aistarterpackage.com