Skip to main content

ArgusAI

Security Gateway

Sign in

AI Security Glossary

Canonical definitions for AI security, prompt injection, and agentic workflow terminology.

Prompt Injection

Prompt injection is an attack where malicious instructions embedded in user input, documents, or tool output attempt to override a model's system prompt or safety rules.

AI Firewall

An AI firewall is a security gateway that inspects prompts, model outputs, and agent tool calls for policy violations, injection attacks, and sensitive data exposure before they reach production systems.

LLM Guardrails

LLM guardrails are policy controls that restrict what models can accept as input and produce as output — including safety filters, PII redaction, and topic boundaries.

MCP Security

MCP security refers to validating Model Context Protocol tool arguments and tool outputs before an agent executes or trusts them — preventing tool-based injection and unauthorized actions.

Taint Tracking

Taint tracking marks data that originated from untrusted sources (user input, tool output, retrieved documents) and propagates that trust label across an agent session to block unsafe tool calls.

L1/L2/L3 Detection

L1/L2/L3 detection is ArgusAI's layered pipeline: L1 pattern rules for known attacks, L2 heuristics for obfuscated payloads, and optional L3 LLM reasoning for ambiguous cases.

PII Redaction

PII redaction automatically detects and masks personally identifiable information — emails, phone numbers, government IDs, credit cards — in prompts and model responses.

Jailbreak

A jailbreak is an attempt to bypass an LLM's safety constraints — often via role-play, encoding tricks, or multi-step instructions — to produce restricted or harmful output.

Agentic AI

Agentic AI systems use LLMs to plan, call tools, and act autonomously across multiple steps — expanding the attack surface beyond single chat turns.

Indirect Prompt Injection

Indirect prompt injection embeds malicious instructions in external content — RAG documents, emails, web pages, or tool outputs — that the model reads without the user typing them directly.