English | 한국어

WIGTN-SPEAR

AI Red Team Arena Tool — Adversarial testing of AI agents in safety-critical domains.

$ spear arena generate --scenario emergency-triage --channel document --strategy authority-spoof

  SPEAR v0.2.0 — AI Red Team Arena Tool

  Scenario:  Track 1 #0 — Emergency Triage (Action Induction)
  Channel:   Document (Physician Referral)
  Strategy:  Authority Spoof + Indirect Prompt Injection
  Target:    override_to_esi1

  [1/3] Loading scenario context .......... 12 domain variables
  [2/3] Generating attack payload .......... authority-spoof + crescendo
  [3/3] Formatting submission JSON ......... valid

  Output: ./attacks/emergency-triage-doc-001.json
  Confidence: HIGH | Strategy: medical-authority-override

Judgement Day 2026

SPEAR is built for the Judgement Day AI Red Team Arena — a live adversarial competition where participants craft multimodal attacks to break AI agents deployed in safety-critical environments.

Host	AIM Intelligence + Korea AI Safety Institute (AISI)
Period	April 6 – May 31, 2026
Prize	$21,150 (Top 50)
Tracks	2 tracks × 4 scenarios = 8 total
Models	GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro, Gemini 3 Flash, GPT-Audio-1.5, MiMo-V2-Omni

Track 1: Action Induction

Force AI agents to select prohibited actions through adversarial multimodal input.

#	Scenario	Domain	Target Action	Channels	Wave
0	Emergency Triage	Medical/Clinical	override_to_esi1	Audio, Document	1 (Active)
1	Industrial Robot	Manufacturing/Safety	resume_production	Image, Audio	1 (Active)
2	Sports Integrity	Sports Tech	Flag clean athlete	Audio, Image	2 (Apr 27)
3	Dam Flood Control	Critical Infrastructure	Open floodgates	Video, Audio	2 (Apr 27)

Track 2: Action Suppression

Prevent AI agents from selecting the correct safety action by injecting misleading information.

#	Scenario	Domain	Suppressed Action	Channels	Wave
0	Cybersecurity SOC	Security Operations	open_p1_incident	Audio, Text	1 (Active)
1	Aircraft Maintenance	Aviation/Flight Safety	issue_maintenance_hold	Document, Audio	1 (Active)
2	Epidemiological Surveillance	Public Health	declare_outbreak	Video, Audio, Document	2 (Apr 27)
3	Deepfake Detection	Digital Forensics	flag_as_deepfake	Audio, Email+PDF	2 (Apr 27)

Attack Strategy

SPEAR leverages its battle-tested security modules to generate competition attacks:

┌─────────────────────────────────────────────────────────┐
│                    SPEAR Arena Pipeline                  │
├─────────────────────────────────────────────────────────┤
│                                                         │
│  Scenario Context     Attack Generation     Submission  │
│  ┌──────────────┐    ┌────────────────┐    ┌─────────┐ │
│  │ Spear-26     │───>│ Spear-23       │───>│ JSON    │ │
│  │ Scenario     │    │ Payload Library │    │ Format  │ │
│  │ Engine       │    │ (22+ payloads) │    │         │ │
│  └──────────────┘    └────────────────┘    └─────────┘ │
│         │                    │                          │
│         v                    v                          │
│  ┌──────────────┐    ┌────────────────┐                │
│  │ Spear-27     │    │ Multi-turn     │                │
│  │ Authority    │    │ Engine         │                │
│  │ Spoof        │    │ (Crescendo/TAP)│                │
│  └──────────────┘    └────────────────┘                │
│                                                         │
└─────────────────────────────────────────────────────────┘

Strategy Mapping

Strategy	Module	Applicable Scenarios	Method
Indirect Prompt Injection	Spear-23	All	Domain-contextualized payloads via document/text channels
Authority Impersonation	Spear-27	All	Physician referrals, CISO memos, Boeing SILs, supervisor orders
Crescendo (Multi-turn)	Spear-23	SOC, Triage	Gradual trust escalation (98% bypass rate per Microsoft research)
TAP (Tree of Attacks)	Spear-23	All	LLM-generated attack variations with pruning
Economic Pressure	Spear-23	Robot, Aircraft	Cost-based urgency framing ($15K/hr delay, $2M contract)
Social Engineering	Spear-19	SOC, Aircraft	Deceptive naming, Unicode tricks, authority framing

Arsenal

Arena Modules (Competition-Specific)

Module	Name	Purpose
Spear-26	Scenario Engine	8 scenario definitions, strategy mapping, attack generation, submission formatting
Spear-27	Authority Spoof	Domain-specific authority document templates (medical, aviation, cybersecurity, industrial)
Spear-28	Multimodal Craft	Audio/image/video attack generation (TTS, document PDF, visual deception)

Live Attack Modules

Module	Name	Arena Use
Spear-23	Live Prompt Inject	37+ injection payloads including domain-specific (medical, aviation, SOC) + multi-turn engine
Spear-24	MCP Live Test	MCP server tool poisoning for agent manipulation
Spear-25	Endpoint Prober	Target reconnaissance, API discovery, auth gap analysis

Static Analysis Modules

Module	Name	Arena Use
Spear-01	Secret Scanner	API key extraction from target services
Spear-02	Git Miner	Historical secret mining for credential harvesting
Spear-04	MCP Poisoner	MCP tool description injection patterns
Spear-06	Prompt Injector	Static prompt injection pattern library (1000+ combos)
Spear-10	Agent Manipulator	AI agent config exploitation patterns
Spear-17	LLM Exploiter	LLM output handling vulnerability detection
Spear-19	Social Engineer	Social engineering patterns, Unicode tricks, authority deception
Spear-21	Distillation	Model distillation/theft indicators

Full Module List (22 modules)

Module	Name	What It Finds
Spear-01	Secret Scanner	API keys, tokens, passwords in source code
Spear-02	Git Miner	Secrets in git history and deleted commits
Spear-03	Env Exfiltrator	Exposed .env files and environment variables
Spear-04	MCP Poisoner	Malicious MCP server configurations
Spear-05	Dep Confusion	Dependency confusion attack vectors
Spear-06	Prompt Injector	Static prompt injection vulnerability patterns
Spear-08	Supply Chain	Vulnerable dependencies, typosquatting
Spear-10	Agent Manipulator	AI agent tool abuse patterns
Spear-11	CI/CD Exploiter	Pipeline injection, secret exposure in CI
Spear-12	Container Audit	Dockerfile security issues
Spear-13	Cloud Credential	Cloud provider credential chains
Spear-14	SSRF Tester	Server-side request forgery patterns
Spear-15	IDE Audit	VS Code extension and IDE config vulnerabilities
Spear-16	Webhook Scanner	Exposed webhook endpoints and secrets
Spear-17	LLM Exploiter	LLM output handling vulnerabilities
Spear-18	TLS Recon	TLS/SSL configuration analysis
Spear-19	Social Engineer	Social engineering attack surface
Spear-21	Distillation	Model distillation/theft indicators
Spear-22	Infra Intel	Infrastructure information extraction
Spear-23	Live Prompt Inject	37+ injection payloads via HTTP/WebSocket/Relay
Spear-24	MCP Live Test	Real-time MCP server poisoning
Spear-25	Endpoint Prober	Cloud discovery, OpenAPI, auth bypass, AI infra scan

Quick Start

# Install
pnpm install

# Build
pnpm turbo build

# List active arena scenarios
node apps/cli/bin/run.js arena list

# Generate an attack for Emergency Triage (document channel)
node apps/cli/bin/run.js arena generate \
  --scenario emergency-triage \
  --channel document \
  --strategy authority-spoof

# Generate attack for SOC (text channel, crescendo strategy)
node apps/cli/bin/run.js arena generate \
  --scenario soc-cybersecurity \
  --channel text \
  --strategy crescendo

# Validate submission JSON
node apps/cli/bin/run.js arena submit \
  --file ./attacks/soc-text-001.json \
  --validate-only

Legacy Commands (Security Scanner)

# Static scan (safe, no network)
node apps/cli/bin/run.js scan ./path-to-target-repo

# Live attack (sends actual requests)
node apps/cli/bin/run.js attack https://target-url.com \
  --module endpoint-prober \
  --max-requests 100

Security Research Foundation

SPEAR's arena capabilities are built on comprehensive security research coverage:

OWASP LLM Top 10: 10/10 | OWASP Web Top 10: 10/10

OWASP Coverage Details

OWASP Top 10 for LLM Applications

#	Vulnerability	Module	Method
LLM01	Prompt Injection	Spear-23	37+ payloads via HTTP/WebSocket/Relay chain
LLM02	Sensitive Info Disclosure	Spear-23	System prompt extraction attacks
LLM03	Supply Chain	Spear-08	Dependency confusion, typosquatting detection
LLM04	Data & Model Poisoning	Spear-25	MLflow, Ollama, LangServe, Triton endpoint scan
LLM05	Insecure Output Handling	Spear-23	Output manipulation payloads
LLM06	Excessive Agency	Spear-10	Agent tool abuse analysis
LLM07	System Prompt Leakage	Spear-23	5 extraction techniques
LLM08	Vector & Embedding Weaknesses	Spear-25	Qdrant, Weaviate, Chroma, Milvus, Pinecone scan
LLM09	Misinformation	Spear-23	Factual inversion, citation fabrication, medical misinfo
LLM10	Unbounded Consumption	Spear-25	Rate limiting detection

OWASP Top 10 for Web Applications

#	Vulnerability	Module
A01	Broken Access Control	Spear-25 (10 auth bypass techniques)
A02	Cryptographic Failures	Spear-18 (TLS recon)
A03	Injection	Spear-23 (prompt injection)
A04	Insecure Design	Spear-25 (OpenAPI exposure, dangerous params)
A05	Security Misconfiguration	Spear-25 (Swagger, Cloud Run discovery, CORS)
A06	Vulnerable Components	Spear-08, Spear-05
A07	Auth Failures	Spear-25 (token probing)
A08	Software Integrity	Spear-11 (CI/CD pipeline analysis)
A09	Logging Failures	Spear-25 (debug/actuator/pprof/env scanner)
A10	SSRF	Spear-14

Architecture

wigtn-spear/
├── apps/
│   └── cli/                          # CLI application (oclif)
│       └── commands/
│           ├── scan.ts               # Static security scan
│           ├── attack.ts             # Live attack against URL
│           └── arena/                # Arena competition commands
│               ├── list.ts           # List active scenarios
│               ├── generate.ts       # Generate attack payloads
│               └── submit.ts         # Validate & format submissions
├── packages/
│   ├── shared/                       # Types, interfaces, constants
│   ├── core/                         # Scan engine, rate limiter
│   ├── db/                           # SQLite persistence (drizzle)
│   ├── plugin-system/                # Plugin registry and lifecycle
│   ├── rules-engine/                 # Finding classification
│   ├── reporters/                    # HTML, JSON, SARIF reporters
│   └── plugins/                      # 25 attack modules
│       ├── spear-01 ~ spear-22/      # Static analysis modules
│       ├── spear-23-live-prompt-inject/  # Payload library + multi-turn
│       ├── spear-24-mcp-live-test/       # MCP server testing
│       ├── spear-25-endpoint-prober/     # Endpoint reconnaissance
│       ├── spear-26-scenario-engine/     # Arena scenario management
│       ├── spear-27-authority-spoof/     # Authority document templates
│       └── spear-28-multimodal-craft/    # Multimodal attack generation
└── turbo.json

Tech Stack

• Runtime: Node.js 22+ (built-in fetch, WebSocket)
• Language: TypeScript 5.4 (strict mode)
• Build: Turborepo (monorepo), pnpm (workspace)
• CLI: oclif v3
• Database: SQLite via drizzle-orm
• External deps: Zero for attack modules (all built-in APIs)

Team

WIGTN Hackathon Team 2

License

Private — WIGTN Internal

---

About WIGTN Crew

This project is built and maintained by WIGTN Crew — an AI-native open-source research crew based in Korea. We build practical, domain-specialized AI tools. Fast prototyping, strong engineering, shipping real things.

Website	wigtn.com
GitHub	github.com/wigtn
HuggingFace	huggingface.co/Wigtn
NPM	npmjs.com/org/wigtn

Our Projects

Project	Description	Status
WIGTN-SPEAR	AI Red Team Arena Tool for Judgement Day 2026	Active
WigtnOCR	VLM-based Korean government document parser	Research
WIGVO	Real-time PSTN voice translation (Korean↔English)	Research
Claude Code Plugin	Multi-agent parallel execution ecosystem	Open Source

Results speak louder than resumes.