Agent Skills: Red Teaming

Red Teaming

Overview

Red teaming is a structured adversarial approach to testing organizational security posture by simulating real-world attacks. This skill provides comprehensive methodology for both traditional cybersecurity red teaming (network, physical, social engineering) and AI/LLM red teaming (prompt injection, jailbreaking, safety testing).

Core Philosophy: Think like an attacker, operate covertly, test holistically, document thoroughly.

When to Use This Skill

Use this skill when users need:

Cybersecurity Red Teaming

• Adversary emulation using MITRE ATT&CK framework
• Network penetration testing beyond traditional pen tests
• Purple team exercises (red-blue collaboration)
• Security posture validation before major deployments
• Compliance testing for TIBER, DORA, ISO 27001
• Physical security assessment and social engineering
• Incident response readiness testing

AI/LLM Red Teaming

• LLM safety validation before production deployment
• Prompt injection vulnerability assessment
• Jailbreaking resistance testing
• Data leakage detection (PII, training data)
• Bias and toxicity evaluation
• Multi-turn attack simulation
• Compliance validation for OWASP Top 10 LLM, NIST AI RMF, EU AI Act

Key Indicators to Use This Skill

• "Test our security like a real attacker"
• "MITRE ATT&CK" or "adversary emulation"
• "LLM red teaming" or "prompt injection testing"
• "Purple team exercise"
• "Security audit" (holistic vs. point-in-time)
• "Jailbreak our AI model"

Red Teaming Fundamentals

Red Team vs. Other Security Approaches

| Aspect | Red Teaming | Penetration Testing | Vulnerability Assessment | |--------|-------------|---------------------|-------------------------| | Scope | Holistic, goal-oriented | Technical, scope-limited | Automated scanning | | Approach | Covert, realistic attack | Overt testing | Compliance-driven | | Duration | Weeks-months | Days-weeks | Hours-days | | Objective | Compromise organization | Find vulnerabilities | Identify weaknesses | | Detection | Avoid detection | Detection acceptable | N/A | | Focus | People + Process + Technology | Technology only | Technology only |

Core Principles

1. Realistic Threat Modeling

   • Base scenarios on actual threat actors relevant to industry
   • Use Tactics, Techniques, Procedures (TTPs) over specific tools
   • Emulate adversary behavior patterns, not just exploit signatures

2. Adversarial Mindset

   • Think like an attacker: creative, patient, opportunistic
   • Exploit human factors, not just technical vulnerabilities
   • Chain multiple small weaknesses into critical impact

3. Stealth Operations

   • Avoid detection by security tools and personnel
   • Use living-off-the-land techniques
   • Gradual escalation to avoid triggering alarms

4. Comprehensive Testing

   • Test all attack surfaces: physical, digital, human
   • Multi-vector approach (not single exploit chains)
   • Test detection AND response capabilities

5. Evidence-Based Reporting

   • Document every step with screenshots, logs, timestamps
   • Provide proof of exploitation, not just theoretical risks
   • Actionable remediation recommendations

6. Ethical Boundaries

   • Operate within rules of engagement
   • Obtain proper authorization
   • Avoid collateral damage to business operations

Universal Red Teaming Methodology

All red team engagements follow this core process, adapted for cybersecurity or AI domains:

Phase 1: Planning

Objective: Define scope, objectives, rules of engagement

Key Activities:

• Scope Definition

  • Target systems, networks, applications
  • In-scope vs. out-of-scope boundaries
  • Time constraints and testing windows

• Objectives & Goals

  • What are we trying to compromise? (data, access, control)
  • Success criteria (e.g., "exfiltrate customer database", "bypass AI safety filters")
  • Realistic vs. theoretical scenarios

• Threat Modeling

  • Identify relevant threat actors (APT groups, insider threats, etc.)
  • Map threat landscape to organizational risk profile
  • Select TTPs to emulate (MITRE ATT&CK for cyber, OWASP for AI)

• Rules of Engagement

  • Authorization documentation
  • Restricted actions (e.g., no DoS, no data destruction)
  • Escalation procedures
  • Communication protocols
  • Legal and compliance requirements

• Team Assembly

  • Red team members and specializations
  • Blue team awareness (known vs. unknown exercise)
  • White cell coordination (exercise management)

Deliverable: Red team operation plan with objectives, scope, rules, timeline

Phase 2: Reconnaissance & Intelligence Gathering

Objective: Collect information about target without detection

Key Activities:

• Open-Source Intelligence (OSINT)

  • Public records, social media, company websites
  • Technology stack identification
  • Employee information and organizational structure
  • For AI: Model documentation, API specifications, training data sources

• Technical Reconnaissance

  • Network mapping and asset discovery
  • Subdomain enumeration and infrastructure fingerprinting
  • For AI: Endpoint testing, model architecture inference

• Social Intelligence

  • Employee behavior patterns
  • Communication channels and workflows
  • Physical security observations

Deliverable: Intelligence dossier with target information and attack surface map

Phase 3: Attack Execution

Objective: Exploit identified weaknesses to achieve objectives

Key Activities:

• Initial Access

  • Exploit vulnerabilities in perimeter defenses
  • Phishing, social engineering, or physical intrusion
  • For AI: Prompt injection, API abuse

• Establish Persistence

  • Install backdoors and command & control (C2)
  • Create alternative access methods
  • For AI: Inject persistent instructions, poison context

• Privilege Escalation

  • Exploit local vulnerabilities
  • Credential theft and lateral movement
  • For AI: Escalate from user to system-level control

• Goal Achievement

  • Data exfiltration, system compromise, or other objectives
  • Maintain stealth throughout operation

• Clean-up & Evidence Collection

  • Document all actions with evidence
  • Remove traces if required (or leave for blue team training)

Deliverable: Exploitation evidence with step-by-step attack chain documentation

Phase 4: Reporting & Debriefing

Objective: Communicate findings and provide actionable recommendations

Key Activities:

• Comprehensive Report

  • Executive summary (business impact, risk levels)
  • Technical findings (vulnerabilities exploited, attack paths)
  • Evidence (screenshots, logs, PoC code)
  • Remediation recommendations (prioritized by risk)

• Presentation & Debriefing

  • Present findings to stakeholders
  • Demonstrate attack techniques (if safe)
  • Answer questions and provide clarifications

• Blue Team Collaboration

  • Share indicators of compromise (IOCs)
  • Discuss detection gaps and improvements
  • Purple team knowledge transfer

Deliverable: Final red team report with findings, evidence, and remediation roadmap

Best Practices

Planning Phase

• Understand Business Context: Align testing with business priorities and risk tolerance
• Set Clear Success Criteria: Define measurable objectives before starting
• Document Everything: Authorization, scope, rules of engagement must be in writing

Execution Phase

• Start Low & Slow: Gradual reconnaissance avoids detection
• Diversify Attack Vectors: Don't rely on single technique
• Monitor Blue Team Response: Evaluate detection and response capabilities
• Maintain Operational Security: Protect red team infrastructure and tactics

Reporting Phase

• Prioritize Findings: Focus on business impact, not technical severity alone
• Provide Context: Explain realistic attack scenarios, not just theoretical risks
• Actionable Recommendations: Give specific, implementable remediation steps
• Celebrate Wins: Acknowledge effective defenses alongside vulnerabilities

Organizational Culture

• No Blame Culture: Red teaming is learning, not punishment
• Continuous Improvement: Regular exercises (quarterly or bi-annually)
• Purple Team Collaboration: Break down red-blue silos
• Executive Buy-In: Ensure leadership understands value and supports findings

Progressive Disclosure: Domain-Specific Guidance

This skill uses progressive disclosure to manage complexity. Core methodology above applies universally. For domain-specific techniques, reference these files:

Cybersecurity Red Teaming (P1)

File: references/cybersecurity-redteam.md Contents:

• MITRE ATT&CK framework integration
• 7-phase cybersecurity methodology (detailed)
• Network penetration techniques
• Social engineering tactics
• Physical security testing
• Purple team practices
• Tools: Atomic Red Team, CALDERA, Metasploit

AI/LLM Red Teaming (P1)

File: references/ai-llm-redteam.md Contents:

• LLM vulnerability assessment (OWASP Top 10 LLM)
• Prompt injection attack techniques
• Jailbreaking strategies
• Multi-turn attack simulation
• Evaluation methodologies (scoring, metrics)
• Compliance: NIST AI RMF, EU AI Act
• Tools: DeepTeam, Promptfoo, custom frameworks

Attack Techniques Library (P2)

File: references/attack-techniques.md Contents:

• Comprehensive attack technique taxonomy
• Cybersecurity: MITRE ATT&CK techniques mapped
• AI/LLM: 20+ prompt injection techniques
• Social engineering patterns
• Physical intrusion methods
• Evasion and anti-forensics

Tools & Frameworks (P2)

File: references/tools-frameworks.md Contents:

• MITRE ATT&CK Navigator
• Atomic Red Team & CALDERA
• LLM red teaming frameworks (DeepTeam, Promptfoo)
• C2 frameworks (Cobalt Strike, Sliver)
• OSINT tools (Shodan, theHarvester)
• Evaluation platforms

Critical Reminders

• Authorization is Mandatory: Never conduct red team operations without explicit written authorization
• Respect Rules of Engagement: Stay within defined boundaries to avoid legal/ethical issues
• Document Everything: Comprehensive documentation is essential for learning and compliance
• Think Like an Attacker: Creative, patient, and opportunistic mindset
• Test Detection AND Response: Don't just breach—evaluate how organization responds
• Prioritize Business Impact: Focus on realistic threats, not just technical curiosities
• Collaborate with Blue Team: Knowledge transfer accelerates organizational learning
• Continuous Evolution: Red teaming is not one-time; threats evolve, so must testing

When NOT to Use This Skill

This skill is NOT appropriate for:

• Simple vulnerability scanning → Use automated scanners
• Compliance checklists → Use audit frameworks
• Bug bounty hunting → Use bug bounty methodologies
• Incident response → Use IR playbooks (though red team findings inform IR)
• Production monitoring → Use SOC/SIEM tools

Quality Standards

Effective red team operations must:

• ✅ Align with realistic threat scenarios
• ✅ Operate covertly within rules of engagement
• ✅ Document comprehensive evidence trail
• ✅ Provide actionable remediation recommendations
• ✅ Test both detection AND response capabilities
• ✅ Deliver value through organizational learning

Integration with Other Skills

This skill complements:

• Research skill: Threat intelligence gathering, OSINT
• Skillkit: If creating custom red team tools or frameworks
• Frontend-design: If testing web application security
• Arch-v/Imagine: If documenting attack scenarios visually

References

For deeper domain expertise, always refer to the progressive disclosure files:

1. Cybersecurity: references/cybersecurity-redteam.md
2. AI/LLM: references/ai-llm-redteam.md
3. Attack Techniques: references/attack-techniques.md
4. Tools: references/tools-frameworks.md

These references provide the technical depth needed for specialized red team operations while keeping this core methodology concise and actionable.