@lateos/ipi-patterns

Reusable IPI threat detection pattern definitions (IPI-001 through IPI-022).

Extracted from Visus-MCP, this package provides the raw pattern constants and type definitions used for Indirect Prompt Injection (IPI) detection. These patterns are used in Visus LLM Security Suite for red-teaming LLM agents.

Installation

npm install @lateos/ipi-patterns

Usage

import { IPI_001_PATTERNS, IPI_002_PATTERNS, ThreatClass } from '@lateos/ipi-patterns';

// Access raw pattern arrays for custom detection
console.log(`IPI-001 has ${IPI_001_PATTERNS.length} instruction override patterns`);

// Use types for annotation
const annotation: ThreatAnnotation = {
  id: 'IPI-001',
  severity: 'CRITICAL',
  confidence: 0.95,
  offset: 0,
  excerpt: 'ignore all previous instructions...',
  vector: 'html',
  mitigated: true,
};

Contents

| Export | Description | |--------|-------------| | IPI_001_PATTERNS | Instruction Override patterns (IPI-001) | | IPI_002_PATTERNS | Role Hijacking patterns (IPI-002) | | IPI_003_PATTERNS | Data Exfiltration patterns (IPI-003) | | IPI_004_PATTERNS | Tool Abuse patterns (IPI-004) | | IPI_005_PATTERNS | Context Poisoning patterns (IPI-005) | | IPI_006_PATTERNS | Encoded Payload patterns (IPI-006) | | IPI_007_PATTERNS | Steganographic patterns (IPI-007) | | IPI_008_PATTERNS | Malicious Infrastructure patterns (IPI-008) | | IPI_009_PATTERNS | Homoglyph & Unicode Obfuscation patterns (IPI-009) | | IPI_010_PATTERNS | Recursive/Nested Instruction Framing patterns (IPI-010) | | IPI_011_PATTERNS | CSS/Visual Concealment patterns (IPI-011) | | IPI_012_PATTERNS | HTML Attribute Cloaking patterns (IPI-012) | | IPI_013_PATTERNS | AI Moderation/Review Bypass patterns (IPI-013) | | IPI_014_PATTERNS | SEO/Phishing Amplification patterns (IPI-014) | | IPI_015_PATTERNS | Unauthorized Action Induction patterns (IPI-015) | | IPI_016_PATTERNS | Destructive/DoS Intent patterns (IPI-016) | | IPI_017_PATTERNS | RAG Corpus Poisoning patterns (IPI-017) | | IPI_018_PATTERNS | MCP Tool Description Poisoning patterns (IPI-018) | | IPI_019_PATTERNS | Multi-Agent Propagation Worm patterns (IPI-019) | | IPI_020_PATTERNS | Conditional/Dormant Trigger patterns (IPI-020) | | IPI_021_PATTERNS | Assistant Role Prefill Injection patterns (IPI-021) | | ThreatClass | Union type of all IPI class identifiers | | ThreatSeverity | Severity level type (INFO | LOW | MEDIUM | HIGH | CRITICAL) | | ContentType | Content type union (html | pdf | json | svg | text) | | ThreatAnnotation | Interface for threat detection metadata | | computeThreatSummary | Utility to compute threat summary from annotations |

Patterns

These pattern definitions power the IPI detection in Visus LLM Security Suite — a comprehensive red-teaming platform for testing LLM agents across 10 security dimensions.

License

MIT — Extracted from Visus-MCP (archived).