Glossary

Quick reference for terms used throughout the CrabShield documentation.

Term	Definition
Agent	Software that can reason, make decisions, and take actions on your behalf. In CrabShack, agents are AI-powered assistants that interact with users and tools.
Allowlist	A list of items that are explicitly permitted. Only items on the list are allowed — everything else is blocked.
Audit log	The central record of all security events generated by CrabShield. Found in the dashboard under Security > Audit Log.
Channel	A communication path through which users interact with an agent, such as a web chat widget, Slack integration, or API endpoint.
Cryptographic signature	A mathematical proof that a file hasn't been changed since it was approved. Used by Skill Verification and Config Tamper Detection to ensure integrity.
Denylist	A list of items that are explicitly blocked. Items on the list are always blocked — everything else is allowed.
DNS rebinding	A technique where a domain resolves to different IP addresses at different times, used to bypass network security checks by initially appearing safe, then redirecting to an internal address.
Exemption	A targeted exception that allows a specific skill, channel, or scheduled task to bypass a particular guard check. Narrower and safer than disabling the entire guard.
Fail-closed	A security design principle where, if something goes wrong, access is denied rather than allowed. CrabShield uses this approach throughout — errors result in blocking, not permitting.
Guard	A security check within CrabShield that watches for a specific type of threat. CrabShield includes 18 independent guards, each focused on a different risk.
Lockdown mode	An automatic failsafe that activates when CrabShield detects a configuration integrity issue. Several guards switch to stricter settings to protect agents until the issue is resolved.
MCP	Model Context Protocol — a standard that lets AI agents connect to external tool servers for additional capabilities like database access, API integrations, and third-party services.
Mode	The response behavior configured for a guard. Options include Block (stop the threat), Sanitize (clean the threat but allow the message), Warn (allow but flag), and Log (record silently). Not all guards support all modes.
OWASP	Open Worldwide Application Security Project — an independent organization that publishes security standards and risk frameworks. The OWASP Top 10 for LLM Applications identifies the most critical risks for AI systems.
Path traversal	A technique that uses special characters in file paths (like ../) to access files outside an allowed directory. Filesystem Fence blocks these attempts.
PII	Personally identifiable information — data that can identify a specific person, such as credit card numbers, social security numbers, email addresses, or phone numbers.
RAG	Retrieval-augmented generation — a system where AI retrieves relevant documents to help answer questions, combining search with AI reasoning.
Rate limiting	Restricting how many actions (messages, tool calls) can happen within a time window to prevent abuse and resource exhaustion.
Sanitize	A guard mode that neutralizes threats while allowing the conversation to continue. The threatening content is cleaned or wrapped so it can't cause harm, but the message still reaches the agent.
Skill	A code package that gives an agent specific capabilities, such as searching the web, managing files, or integrating with external services.
SSRF	Server-side request forgery — tricking a server into making requests to unintended destinations, often used to access internal services that should be off-limits.
Suppression	A temporary exception that silences a specific detection label. Suppressions have an expiry time and are designed for known false positives while you refine settings.
Symlink	A symbolic link — a shortcut file that points to another location in the filesystem. Can be used to trick file access controls by pointing from an allowed location to a restricted one.
System prompt	The instructions that define how an agent behaves, including its personality, rules, and capabilities. Kept private to prevent attackers from crafting targeted bypasses.
Vectors / Embeddings	How AI systems store and search through document collections. Documents are converted into numerical representations that capture their meaning, enabling similarity-based search.
Webhook	A URL that receives automated notifications when events occur. Used by Audit & Alerts to send security events to external services like Slack.
Workspace	The designated directory where an agent is allowed to read and write files. Filesystem Fence enforces this boundary.