CrabShield Security Overview

CrabShield is the built-in security layer for CrabShack. It monitors every conversation and action your AI agents take, detecting and preventing threats in real time — no setup required.

Defense in Depth

CrabShield uses 18 independent security guards, each focused on a different type of threat. Because they work independently, a problem with one guard never compromises the others. And if anything unexpected happens, CrabShield automatically switches to its strictest settings rather than leaving you unprotected.

How It Works

CrabShield watches your agents at every stage:

• Incoming messages are checked for manipulation attempts, unauthorized access, and sensitive data
• Agent actions like tool usage, file access, and network requests are validated against your security policies
• Agent responses are scanned for potentially dangerous content and data leaks
• System integrity is continuously monitored to ensure nothing has been tampered with

When a threat is detected, CrabShield takes action based on your settings — blocking it, cleaning it up, or flagging it for review.

Guards at a Glance

Guard	What It Protects Against
Prompt Injection	Attempts to hijack or manipulate your agent's behavior
System Prompt Leakage	Attempts to extract your agent's instructions
Tool Gate	Unauthorized or dangerous tool usage
Filesystem Fence	Unauthorized file access outside allowed areas
Credential Vault	API keys, passwords, and personal information appearing in conversations
Escalation Detection	Commands that try to gain elevated system access
Input Validator	Oversized messages and message flooding
Output Scanner	Dangerous code patterns in AI-generated responses
Memory Guard	Tampering with agent memory and conversation context
MCP Hardening	Unauthorized external tool server connections
Network Policy	Unauthorized network requests to internal or sensitive destinations
Skill Verification	Tampered or unauthorized agent skills
Config Tamper Detection	Unauthorized changes to critical configuration files
Auth Enforcement	Messages from unauthorized channels or users
Audit & Alerts	Complete security event logging and real-time alerts
RAG Scanner	Hidden manipulation attempts in uploaded Knowledge Base documents
Anomaly Detection	Unusual agent behavior compared to its normal pattern
Exemption Engine	Not a threat detector — lets you exempt known-safe skills, jobs, or channels from other guards
Policy Compliance	Gaps or accidental weakening of your own CrabShield settings

Guards by Category

Not every guard is relevant to every user. Here's how they break down:

Core Protection

These guards are relevant to everyone using CrabShack:

• Prompt Injection — stops attempts to manipulate your agent's behavior
• System Prompt Leakage — prevents extraction of your agent's instructions
• Credential Vault — redacts API keys, passwords, and personal information
• Input Validator — enforces message size limits and rate limiting
• Output Scanner — catches dangerous code patterns in agent responses
• Escalation Detection — detects commands that try to gain elevated access
• RAG Scanner — scans uploaded Knowledge Base documents for hidden manipulation attempts

Infrastructure Controls

Primarily relevant to administrators and self-hosters:

• Tool Gate — controls which tools your agent can use
• Filesystem Fence — restricts file access to authorized areas
• Network Policy — controls where your agent can make network requests
• Memory Guard — protects against poisoned memories and session limits
• Auth Enforcement — ensures messages come from authorized channels

Advanced / Developer

Primarily relevant to developers and teams using MCP servers or custom skills:

• MCP Hardening — controls external tool server connections
• Skill Verification — ensures agent skills haven't been tampered with
• Config Tamper Detection — monitors critical files for unauthorized changes
• Exemption Engine — lets you exempt specific skills, scheduled jobs, or channels from other guards

Operational

Relevant to everyone:

• Audit & Alerts — complete security event logging and real-time alerts
• Anomaly Detection — flags unusual agent behavior compared to its own normal pattern
• Policy Compliance — checks that your own security settings haven't drifted or been weakened

Secure by Default

CrabShield is active from the moment you start using CrabShack. All guards are enabled with balanced defaults that protect against the most common threats without getting in the way of normal use. If you need more or less protection, you can adjust everything from the dashboard — see the Configuration page to learn how.