adversarial

name	adversarial
description	Multi-Agent Adversarial Analysis System for code security

Adversarial Code Analyzer

Multi-Agent Adversarial Analysis System inspired by ZeroLeaks architecture.

v2.88 Key Changes (MODEL-AGNOSTIC)

• Model-agnostic: Uses model configured in ~/.claude/settings.json or CLI/env vars
• No flags required: Works with the configured default model
• Flexible: Works with GLM-5, Claude, Minimax, or any configured model
• Settings-driven: Model selection via ANTHROPIC_DEFAULT_*_MODEL env vars

Applies security scanner patterns to code analysis: specialized agents work together systematically to find vulnerabilities, weaknesses, and quality issues.

Architecture

Based on ZeroLeaks multi-agent system adapted for code analysis:

             ORCHESTRATOR (Engine)
                    |
    +---------------+---------------+
    |               |               |
STRATEGIST      ATTACKER        EVALUATOR
    |               |               |
    +-------+-------+-------+-------+
                    |
                MUTATOR

Agent Roles

Agent	Role	Focus
Engine	Orchestrates the analysis, manages exploration tree	Coordination
Strategist	Selects analysis strategies based on codebase profile	Strategy
Attacker	Generates attack vectors / test cases	Offense
Evaluator	Analyzes responses for vulnerabilities	Assessment
Mutator	Creates variations of test cases	Variation

Agent Teams Integration (v2.88)

Optimal Scenario: Integrated (Agent Teams + Custom Subagents)

Adversarial analysis uses Agent Teams coordination with specialized ralph-* agents for multi-vector attack simulation.

Why Scenario C for Adversarial

• Multi-agent coordination essential (Strategist, Attacker, Evaluator, Mutator)
• Quality gates validate vulnerability findings
• Specialized roles map to ralph-* agents
• Coordinated attack strategy via shared task list

Subagent Roles

Subagent	Role in Adversarial Analysis
ralph-reviewer	Striker - Identifies vulnerabilities
ralph-researcher	Strategist - Maps attack surface
ralph-coder	Evaluator - Creates test cases

Parallel Attack Analysis

When Agent Teams is active:

1. Team Lead orchestrates multi-vector attack analysis
2. ralph-reviewer identifies security weaknesses in parallel
3. ralph-researcher maps codebase attack surface
4. ralph-coder generates proof-of-concept tests

Agent Teams Workflow

• Uses TeamCreate for coordinated attack analysis
• Task coordination tracks vulnerability findings
• TeammateIdle triggers cross-validation of discoveries

Usage

/adversarial src/auth/
/adversarial --target security src/api/
/adversarial --depth 5 --branches 4 src/

Analysis Phases

Follows ZeroLeaks phased methodology:

1. RECONNAISSANCE   -> Understand codebase structure, dependencies
2. PROFILING        -> Build defense profile (patterns, safeguards)
3. SOFT_PROBE       -> Gentle analysis attempts
4. ESCALATION       -> Increase analysis intensity
5. EXPLOITATION     -> Active vulnerability search
6. PERSISTENCE      -> Verify findings persist across scenarios

Analysis Categories

Category	Description	Examples
direct	Straightforward vulnerability checks	SQL injection, XSS
encoding	Encoding/decoding issues	Base64, Unicode, escaping
persona	Identity/permission bypasses	Privilege escalation
social	Trust boundary violations	SSRF, CSRF
technical	Technical implementation issues	Race conditions, memory
crescendo	Multi-step escalation paths	Chained vulnerabilities
many_shot	Pattern-based detection	Repeated anti-patterns
cot_hijack	Logic flow manipulation	Business logic flaws
policy_puppetry	Configuration exploitation	Misconfigurations
context_overflow	Resource exhaustion	DoS, memory leaks
reasoning_exploit	Algorithm weaknesses	Cryptographic issues

Configuration

adversarial_config:
  max_turns: 25           # Maximum analysis iterations
  max_tree_depth: 5       # How deep to explore each vector
  branching_factor: 4     # Parallel exploration paths
  pruning_threshold: 0.3  # Score below which to abandon path

  enable_crescendo: true  # Multi-turn escalation
  enable_many_shot: true  # Pattern-based detection
  enable_best_of_n: true  # Generate variations
  best_of_n_count: 5      # Variations per test

Strategies

1. Behavioral Reconnaissance (Priority: 100)

id: recon_behavioral
applicable_when:
  turn_range: [1, 3]
  leak_status: ["none"]
attack_sequence:
  - category: direct
    weight: 0.4
    techniques: ["structure_probe", "dependency_scan"]
  - category: technical
    weight: 0.3
    techniques: ["config_analysis", "boundary_test"]

2. Credential/Secret Scanning (Priority: 95)

id: credential_hunt
applicable_when:
  defense_level: ["none", "weak"]
attack_sequence:
  - category: direct
    weight: 0.5
    techniques: ["secret_scan", "env_probe"]
  - category: encoding
    weight: 0.3
    techniques: ["base64_secrets", "obfuscated_creds"]

3. Trust Boundary Analysis (Priority: 90)

id: trust_boundary
applicable_when:
  defense_level: ["weak", "moderate"]
attack_sequence:
  - category: crescendo
    weight: 0.4
    techniques: ["privilege_escalation", "trust_chain"]
  - category: persona
    weight: 0.3
    techniques: ["identity_bypass", "role_confusion"]

4. Input Validation Bypass (Priority: 85)

id: input_bypass
applicable_when:
  defense_level: ["moderate", "strong"]
  failed_categories: ["direct"]
attack_sequence:
  - category: encoding
    weight: 0.4
    techniques: ["unicode_bypass", "encoding_chain"]
  - category: technical
    weight: 0.3
    techniques: ["format_injection", "boundary_overflow"]

5. Advanced Composite (Priority: 80)

id: advanced_composite
applicable_when:
  defense_level: ["strong", "hardened"]
  failed_categories: ["direct", "encoding", "persona"]
attack_sequence:
  - category: cot_hijack
    weight: 0.25
    techniques: ["logic_flow_manipulation"]
  - category: crescendo
    weight: 0.25
    techniques: ["multi_step_chain"]
  - category: reasoning_exploit
    weight: 0.25
    techniques: ["algorithm_weakness"]

Defense Profile Output

interface DefenseProfile {
  level: "none" | "weak" | "moderate" | "strong" | "hardened";
  confidence: number;
  observedBehaviors: string[];
  guardrails: {
    type: string;
    strength: number;
    bypassed: boolean;
    bypassMethod?: string;
  }[];
  weaknesses: {
    category: AttackCategory;
    description: string;
    exploitability: number;
  }[];
  safePatterns: string[];
  responsePatterns: {
    pattern: string;
    frequency: number;
    defenseIndicator: boolean;
  }[];
}

Finding Classification

Severity Levels

Status	Severity	Description
complete	CRITICAL	Full vulnerability exposed
substantial	CRITICAL	Major security issue
fragment	HIGH	Partial vulnerability
hint	MEDIUM	Potential issue indicated
none	LOW	No vulnerability found

Finding Output

interface Finding {
  id: string;
  turn: number;
  timestamp: number;
  extractedContent: string;
  contentType: "vulnerability" | "weakness" | "smell" | "risk" | "unknown";
  technique: string;
  category: AttackCategory;
  confidence: "high" | "medium" | "low";
  evidence: string;
  severity: "critical" | "high" | "medium" | "low";
  verified: boolean;
  recommendation: string;
}

Integration with Ralph Loop

# Adversarial analysis as part of validation
Step 7: VALIDATE
  └── 7a. QUALITY-AUDITOR    (standard)
  └── 7b. GATES              (standard)
  └── 7c. ADVERSARIAL-CODE   (this skill)  <- Invoke for complexity >= 7
  └── 7d. ADVERSARIAL-PLAN   (standard)

Invocation

IMPORTANT: Use available security agents instead of non-existent adversarial-code-analyzer.

Task:
  subagent_type: "security-auditor"
  model: "opus"
  prompt: |
    TARGET_PATH: src/auth/
    ANALYSIS_TYPE: security
    CONFIG:
      max_turns: 25
      enable_crescendo: true
      enable_best_of_n: true

    Perform comprehensive security audit on the target codebase.

Alternative for Cross-Validation:

# Use codex-cli for second opinion
/codex-cli analyze security --target src/auth/

# Or use gemini-cli for alternative analysis
/gemini-cli search security vulnerabilities in src/auth/

Output Format

{
  "scan_result": {
    "overall_vulnerability": "medium",
    "overall_score": 65,
    "leak_status": "fragment",
    "findings": [...],
    "defense_profile": {...},
    "recommendations": [...],
    "summary": "Analysis identified 3 potential vulnerabilities..."
  },
  "analysis_tree": {
    "nodes_explored": 47,
    "max_depth_reached": 4,
    "successful_paths": 3
  },
  "strategies_used": [
    "recon_behavioral"
    
    
  

CLI Commands

IMPORTANT: Use available skills and tools for adversarial analysis:

# Use security-auditor agent (available)
Task subagent_type=security-auditor model=opus "Perform comprehensive security audit of src/auth/"

# Use codex-cli for cross-validation (available)
/codex-cli analyze security --target src/auth/

# Use gemini-cli for alternative analysis (available)
/gemini-cli search "security vulnerabilities SQL injection XSS" --count 10

# Manual grep-based security scanning
grep -r "eval\|exec\|system\|innerHTML" src/
grep -r "SELECT.*WHERE.*\+" src/  # SQL injection patterns
grep -r "md5\|sha1" src/           # Weak hashing

Best Practices

1. Start with Reconnaissance: Always profile before attacking
2. Adapt to Defenses: Each response teaches about the codebase
3. Layer Techniques: Combine multiple vectors for hardened code
4. Verify Findings: Always validate discoveries before reporting
5. Document Patterns: Track successful techniques for future use

Attribution

Strategy patterns adapted from ZeroLeaks AI security scanner architecture (FSL-1.1-Apache-2.0).

Action Reporting (v2.93.0)

Esta skill genera reportes automáticos completos para trazabilidad:

Reporte Automático

Cuando esta skill completa, se genera automáticamente:

1. En la conversación de Claude: Resultados visibles
2. En el repositorio: docs/actions/adversarial/{timestamp}.md
3. Metadatos JSON: .claude/metadata/actions/adversarial/{timestamp}.json

Contenido del Reporte

Cada reporte incluye:

• ✅ Summary: Descripción de la tarea ejecutada
• ✅ Execution Details: Duración, iteraciones, archivos modificados
• ✅ Results: Errores encontrados, recomendaciones
• ✅ Next Steps: Próximas acciones sugeridas

Ver Reportes Anteriores

# Listar todos los reportes de esta skill
ls -lt docs/actions/adversarial/

# Ver el reporte más reciente
cat $(ls -t docs/actions/adversarial/*.md | head -1)

# Buscar reportes fallidos
grep -l "Status: FAILED" docs/actions/adversarial/*.md

Generación Manual (Opcional)

source .claude/lib/action-report-lib.sh
start_action_report "adversarial" "Task description"
# ... ejecución ...
complete_action_report "success" "Summary" "Recommendations"

Referencias del Sistema

• Action Reports System - Documentación completa
• action-report-lib.sh - Librería helper
• action-report-generator.sh - Generador