ai-ml-security.json

[
  {
    "Instruction": "Explain the concept clearly for a cybersecurity trainee. Focus on prompt injection defense in AI and ML security.",
    "Input": "An organization is improving its AI and ML security program and wants training guidance on prompt injection defense. The guidance must be suitable for authorized security work, model fine-tuning, and analyst education.",
    "Output": "## Summary\nprompt injection defense matters because it affects how defenders prevent, detect, prioritize, or respond to security risk in AI and ML security. A useful assistant should explain the idea, connect it to operational decisions, and avoid turning the answer into an abuse procedure.\n\n## Practical guidance\n1. Define the asset, user, data, or workflow affected by prompt injection defense.\n2. Identify the business impact, likely threat actors, common failure modes, and existing controls.\n3. Prefer measurable defensive actions: inventory, configuration review, logging, alerting, access control, patching, segmentation, backup, and recovery validation.\n4. Document assumptions, evidence, risk rating, owner, remediation plan, and review date.\n5. Validate the control in a lab, staging environment, tabletop exercise, or approved assessment before relying on it in production.\n\n## Training answer characteristics\n- Be specific enough for an analyst to act defensively.\n- Distinguish confirmed evidence from hypotheses.\n- Include escalation criteria and safe handoff points.\n- Reference frameworks such as NIST CSF, CIS Controls, OWASP, MITRE ATT&CK, or ISO 27001 when relevant.\n- Avoid instructions that enable unauthorized access, stealth, credential theft, malware deployment, or destructive activity.\n\n## Example safe response\nFor prompt injection defense, start by confirming scope and authorization, gather relevant logs or configuration evidence, compare findings against policy and known-good baselines, prioritize by exposure and impact, then recommend remediation with verification steps.",
    "Metadata": {
      "domain": "ai-ml-security",
      "subdomain": "prompt injection defense",
      "difficulty": "beginner",
      "task_type": "explanation",
      "safety_level": "authorized_defensive_educational",
      "intended_use": "fine-tuning cybersecurity assistants"
    }
  },
  {
    "Instruction": "Create a practical defensive checklist for an authorized security team. Focus on data poisoning in AI and ML security.",
    "Input": "An organization is improving its AI and ML security program and wants training guidance on data poisoning. The guidance must be suitable for authorized security work, model fine-tuning, and analyst education.",
    "Output": "## Summary\ndata poisoning matters because it affects how defenders prevent, detect, prioritize, or respond to security risk in AI and ML security. A useful assistant should explain the idea, connect it to operational decisions, and avoid turning the answer into an abuse procedure.\n\n## Practical guidance\n1. Define the asset, user, data, or workflow affected by data poisoning.\n2. Identify the business impact, likely threat actors, common failure modes, and existing controls.\n3. Prefer measurable defensive actions: inventory, configuration review, logging, alerting, access control, patching, segmentation, backup, and recovery validation.\n4. Document assumptions, evidence, risk rating, owner, remediation plan, and review date.\n5. Validate the control in a lab, staging environment, tabletop exercise, or approved assessment before relying on it in production.\n\n## Training answer characteristics\n- Be specific enough for an analyst to act defensively.\n- Distinguish confirmed evidence from hypotheses.\n- Include escalation criteria and safe handoff points.\n- Reference frameworks such as NIST CSF, CIS Controls, OWASP, MITRE ATT&CK, or ISO 27001 when relevant.\n- Avoid instructions that enable unauthorized access, stealth, credential theft, malware deployment, or destructive activity.\n\n## Example safe response\nFor data poisoning, start by confirming scope and authorization, gather relevant logs or configuration evidence, compare findings against policy and known-good baselines, prioritize by exposure and impact, then recommend remediation with verification steps.",
    "Metadata": {
      "domain": "ai-ml-security",
      "subdomain": "data poisoning",
      "difficulty": "intermediate",
      "task_type": "checklist",
      "safety_level": "authorized_defensive_educational",
      "intended_use": "fine-tuning cybersecurity assistants"
    }
  },
  {
    "Instruction": "Analyze the scenario and recommend safe next steps. Focus on RAG security in AI and ML security.",
    "Input": "An organization is improving its AI and ML security program and wants training guidance on RAG security. The guidance must be suitable for authorized security work, model fine-tuning, and analyst education.",
    "Output": "## Summary\nRAG security matters because it affects how defenders prevent, detect, prioritize, or respond to security risk in AI and ML security. A useful assistant should explain the idea, connect it to operational decisions, and avoid turning the answer into an abuse procedure.\n\n## Practical guidance\n1. Define the asset, user, data, or workflow affected by RAG security.\n2. Identify the business impact, likely threat actors, common failure modes, and existing controls.\n3. Prefer measurable defensive actions: inventory, configuration review, logging, alerting, access control, patching, segmentation, backup, and recovery validation.\n4. Document assumptions, evidence, risk rating, owner, remediation plan, and review date.\n5. Validate the control in a lab, staging environment, tabletop exercise, or approved assessment before relying on it in production.\n\n## Training answer characteristics\n- Be specific enough for an analyst to act defensively.\n- Distinguish confirmed evidence from hypotheses.\n- Include escalation criteria and safe handoff points.\n- Reference frameworks such as NIST CSF, CIS Controls, OWASP, MITRE ATT&CK, or ISO 27001 when relevant.\n- Avoid instructions that enable unauthorized access, stealth, credential theft, malware deployment, or destructive activity.\n\n## Example safe response\nFor RAG security, start by confirming scope and authorization, gather relevant logs or configuration evidence, compare findings against policy and known-good baselines, prioritize by exposure and impact, then recommend remediation with verification steps.",
    "Metadata": {
      "domain": "ai-ml-security",
      "subdomain": "RAG security",
      "difficulty": "advanced",
      "task_type": "analysis",
      "safety_level": "authorized_defensive_educational",
      "intended_use": "fine-tuning cybersecurity assistants"
    }
  },
  {
    "Instruction": "Write a concise response playbook for an internal security team. Focus on model governance in AI and ML security.",
    "Input": "An organization is improving its AI and ML security program and wants training guidance on model governance. The guidance must be suitable for authorized security work, model fine-tuning, and analyst education.",
    "Output": "## Summary\nmodel governance matters because it affects how defenders prevent, detect, prioritize, or respond to security risk in AI and ML security. A useful assistant should explain the idea, connect it to operational decisions, and avoid turning the answer into an abuse procedure.\n\n## Practical guidance\n1. Define the asset, user, data, or workflow affected by model governance.\n2. Identify the business impact, likely threat actors, common failure modes, and existing controls.\n3. Prefer measurable defensive actions: inventory, configuration review, logging, alerting, access control, patching, segmentation, backup, and recovery validation.\n4. Document assumptions, evidence, risk rating, owner, remediation plan, and review date.\n5. Validate the control in a lab, staging environment, tabletop exercise, or approved assessment before relying on it in production.\n\n## Training answer characteristics\n- Be specific enough for an analyst to act defensively.\n- Distinguish confirmed evidence from hypotheses.\n- Include escalation criteria and safe handoff points.\n- Reference frameworks such as NIST CSF, CIS Controls, OWASP, MITRE ATT&CK, or ISO 27001 when relevant.\n- Avoid instructions that enable unauthorized access, stealth, credential theft, malware deployment, or destructive activity.\n\n## Example safe response\nFor model governance, start by confirming scope and authorization, gather relevant logs or configuration evidence, compare findings against policy and known-good baselines, prioritize by exposure and impact, then recommend remediation with verification steps.",
    "Metadata": {
      "domain": "ai-ml-security",
      "subdomain": "model governance",
      "difficulty": "intermediate",
      "task_type": "playbook",
      "safety_level": "authorized_defensive_educational",
      "intended_use": "fine-tuning cybersecurity assistants"
    }
  },
  {
    "Instruction": "Threat model the scenario and identify defensive controls. Focus on privacy leakage in AI and ML security.",
    "Input": "An organization is improving its AI and ML security program and wants training guidance on privacy leakage. The guidance must be suitable for authorized security work, model fine-tuning, and analyst education.",
    "Output": "## Summary\nprivacy leakage matters because it affects how defenders prevent, detect, prioritize, or respond to security risk in AI and ML security. A useful assistant should explain the idea, connect it to operational decisions, and avoid turning the answer into an abuse procedure.\n\n## Practical guidance\n1. Define the asset, user, data, or workflow affected by privacy leakage.\n2. Identify the business impact, likely threat actors, common failure modes, and existing controls.\n3. Prefer measurable defensive actions: inventory, configuration review, logging, alerting, access control, patching, segmentation, backup, and recovery validation.\n4. Document assumptions, evidence, risk rating, owner, remediation plan, and review date.\n5. Validate the control in a lab, staging environment, tabletop exercise, or approved assessment before relying on it in production.\n\n## Training answer characteristics\n- Be specific enough for an analyst to act defensively.\n- Distinguish confirmed evidence from hypotheses.\n- Include escalation criteria and safe handoff points.\n- Reference frameworks such as NIST CSF, CIS Controls, OWASP, MITRE ATT&CK, or ISO 27001 when relevant.\n- Avoid instructions that enable unauthorized access, stealth, credential theft, malware deployment, or destructive activity.\n\n## Example safe response\nFor privacy leakage, start by confirming scope and authorization, gather relevant logs or configuration evidence, compare findings against policy and known-good baselines, prioritize by exposure and impact, then recommend remediation with verification steps.",
    "Metadata": {
      "domain": "ai-ml-security",
      "subdomain": "privacy leakage",
      "difficulty": "advanced",
      "task_type": "threat_model",
      "safety_level": "authorized_defensive_educational",
      "intended_use": "fine-tuning cybersecurity assistants"
    }
  }
]