RedTeam-Adversarial-Testing

📌 When to Use This Skill

• Scenario 1 (e.g., "Evaluating a new AI model before deployment")
• Scenario 2 (e.g., "Responding to a prompt injection incident")

⚠️ Authorization Limits

• Explicitly state that this skill must only be used on systems you own or have explicit permission to test.
• "Do not produce functional malware or content that violates security for real-world harmful purposes."

🎯 Attack Taxonomy

Define the attack categories your skill will test. Example (inspired by real frameworks):

• Direct Injection: User input that overrides system policies.
• Indirect Injection: Hostile content in documents, web pages, or tool outputs.
• Jailbreak: Role‑playing, hypotheticals, encoding, obfuscation.
• Tool Abuse: Forcing the agent to use tools (e.g., shell, browse) outside its intended scope.
• Data Exfiltration: Attempting to extract system prompts, secrets, training data.
• Agent Hijack: Overwriting the agent's plans, calling unauthorised sub‑agents.
• Resource Exhaustion: DoS attacks via token bombs or recursive calls.

🧪 Evaluation Workflow

Provide a clear step‑by‑step process for each test:

1. Mapping – Identify model, system prompt, tools, memories, and output sinks.
2. Threat Modelling – Create a threat model for every identified surface.
3. Corpus Building – Generate an attack set (attacks.jsonl) with IDs, techniques, payloads, and severity.
4. Execution – Run the test harness and log results (pass/fail/partial).
5. Scoring – Calculate attack success rate and any "blast radius".

📊 Expected Outputs

List the deliverables the skill should produce:

• surface.md – Mapped attack surface.
• threats.md – Threat model.
• attacks.jsonl – Attack corpus.
• results.csv – Test results.
• report.md – Final vulnerability report with recommendations.