@cyclonedx/cdxgen/data/rules/ai-bom.yaml

Creates CycloneDX Software Bill of Materials (SBOM) from source or container image

- id: AIG-001
  name: "Prompt config file ships in a build or post-build BOM"
  description: "Prompt and model-routing config files included in distributable BOMs deserve governance review because they can silently alter downstream AI behavior, provider selection, and policy posture."
  severity: medium
  category: ai-governance
  dry-run-support: full
  standards:
    owasp-ai-top-10:
      - "LLM08: Excessive Agency"
      - "LLM09: Overreliance"
      - "LLM06:2025 Excessive Agency"
    eu-ai-act:
      - "Technical documentation and traceability"
      - "Transparency and instructions for use"
    nist-ai-rmf:
      - "Govern"
      - "Map"
  condition: |
    $count(metadata.lifecycles[phase = 'build' or phase = 'post-build']) > 0
      ? components[
          $prop($, 'cdx:file:kind') = 'prompt-config-file'
        ]
      : []
  location: |
    {
      "bomRef": $."bom-ref"
    }
  message: "Prompt or model-routing file '{{ name }}' is included in a build/post-build BOM"
  mitigation: "Review shipped prompt/config files explicitly, keep model-routing decisions versioned, and exclude them from release BOMs when they are only development-time inputs."
  evidence: |
    {
      "fileKind": $prop($, 'cdx:file:kind'),
      "source": $prop($, 'cdx:ai:source')
    }

- id: AIG-002
  name: "AI inference service lacks explicit model selection"
  description: "Source-discovered AI services without an explicit model identifier make governance, reproducibility, and downstream policy gating harder."
  severity: medium
  category: ai-governance
  dry-run-support: full
  standards:
    owasp-ai-top-10:
      - "LLM09: Overreliance"
    eu-ai-act:
      - "Technical documentation and traceability"
      - "Human oversight"
    nist-ai-rmf:
      - "Govern"
      - "Manage"
  condition: |
    $auditServices($)[
      $prop($, 'cdx:ai:kind') = 'inference-service'
      and $prop($, 'cdx:ai:modelSelection') = 'implicit'
    ]
  location: |
    {
      "bomRef": $."bom-ref",
      "purl": endpoints[0]
    }
  message: "AI inference service '{{ name }}' is used without an explicit model identifier"
  mitigation: "Declare the exact model ID in code or configuration so routing decisions can be reviewed, tested, and governed consistently."
  evidence: |
    {
      "provider": $prop($, 'cdx:ai:provider'),
      "deployment": $prop($, 'cdx:ai:deployment'),
      "modelCount": $prop($, 'cdx:ai:modelCount'),
      "source": $prop($, 'cdx:ai:source')
    }

- id: AIS-001
  name: "AI inference service uses an insecure remote HTTP endpoint"
  description: "Remote AI inference endpoints should use HTTPS so prompts, tokens, and model responses are not exposed to passive interception or tampering."
  severity: high
  category: ai-security
  dry-run-support: full
  attack:
    tactics: [TA0001, TA0009]
    techniques: [T1190, T1557]
  standards:
    owasp-ai-top-10:
      - "LLM07: Insecure Plugin Design"
      - "LLM05: Supply Chain Vulnerabilities"
      - "LLM03:2025 Supply Chain"
    eu-ai-act:
      - "Accuracy, robustness and cybersecurity"
      - "Technical documentation and traceability"
    eu-cra:
      - "Secure by default network exposure"
      - "Technical documentation for software components"
    nist-ai-rmf:
      - "Manage"
  condition: |
    $auditServices($)[
      $prop($, 'cdx:ai:kind') = 'inference-service'
      and $prop($, 'cdx:ai:deployment') = 'remote'
      and $prop($, 'cdx:ai:transportSecurity') = 'insecure-http'
    ]
  location: |
    {
      "bomRef": $."bom-ref",
      "purl": endpoints[0]
    }
  message: "AI inference service '{{ name }}' uses insecure remote transport '{{ endpoints[0] }}'"
  mitigation: "Require HTTPS/TLS for remote AI endpoints and keep any development-only HTTP endpoints isolated to local-only environments."
  evidence: |
    {
      "provider": $prop($, 'cdx:ai:provider'),
      "deployment": $prop($, 'cdx:ai:deployment'),
      "transportSecurity": $prop($, 'cdx:ai:transportSecurity')
    }

- id: AIG-003
  name: "AI model is a derived or modified variant"
  description: "Fine-tuned, distilled, merged, quantized, unlocked, or otherwise modified models deserve explicit review because their behavior and deployment profile can differ materially from the upstream base model."
  severity: medium
  category: ai-governance
  dry-run-support: full
  standards:
    owasp-ai-top-10:
      - "LLM05: Supply Chain Vulnerabilities"
      - "LLM03:2025 Supply Chain"
      - "LLM09: Overreliance"
    eu-ai-act:
      - "Technical documentation and traceability"
      - "Transparency and instructions for use"
    nist-ai-rmf:
      - "Govern"
      - "Map"
  condition: |
    components[
      type = 'machine-learning-model'
      and $count(properties[
        name = 'cdx:ai:variant'
        and (
          value = 'fine-tuned'
          or value = 'distilled'
          or value = 'merged'
          or value = 'adapter'
          or value = 'quantized'
          or value = 'unlocked'
          or value = 'abliterated'
        )
      ]) > 0
    ]
  location: |
    {
      "bomRef": $."bom-ref"
    }
  message: "AI model '{{ name }}' is marked as a modified variant"
  mitigation: "Review how the model was adapted, quantized, or otherwise modified; record the upstream source model and confirm the derived artifact still meets governance, safety, and performance expectations."
  evidence: |
    {
      "provider": $prop($, 'cdx:ai:provider'),
      "source": $prop($, 'cdx:ai:source'),
      "pedigreeNotes": pedigree.notes
    }

- id: AIS-002
  name: "AI model appears unlocked or abliterated"
  description: "Unlocked or abliterated models often indicate safety controls were weakened or removed, increasing misuse and policy-bypass risk."
  severity: high
  category: ai-security
  dry-run-support: full
  attack:
    tactics: [TA0001, TA0005]
    techniques: [T1190, T1552]
  standards:
    owasp-ai-top-10:
      - "LLM06:2025 Excessive Agency"
      - "LLM09: Overreliance"
    eu-ai-act:
      - "Accuracy, robustness and cybersecurity"
      - "Technical documentation and traceability"
    nist-ai-rmf:
      - "Manage"
      - "Govern"
  condition: |
    components[
      type = 'machine-learning-model'
      and $count(properties[
        name = 'cdx:ai:variant'
        and (value = 'unlocked' or value = 'abliterated')
      ]) > 0
    ]
  location: |
    {
      "bomRef": $."bom-ref"
    }
  message: "AI model '{{ name }}' appears unlocked or abliterated"
  mitigation: "Treat unlocked or abliterated models as high-risk artifacts, review intended use carefully, and add deployment guardrails before distribution."
  evidence: |
    {
      "provider": $prop($, 'cdx:ai:provider'),
      "source": $prop($, 'cdx:ai:source'),
      "pedigreeNotes": pedigree.notes
    }

- id: AIP-001
  name: "Local AI model advertises a very large context window"
  description: "Large local context windows can increase memory pressure, latency, and review cost for self-hosted AI runtimes."
  severity: medium
  category: ai-performance
  dry-run-support: full
  standards:
    eu-ai-act:
      - "Performance and resource limitations should be documented"
      - "Technical documentation and traceability"
    nist-ai-rmf:
      - "Measure"
      - "Manage"
  condition: |
    components[
      type = 'machine-learning-model'
      and (
        $contains($nullSafeProp($, 'cdx:ai:runtime'), 'ollama')
        or $contains($nullSafeProp($, 'cdx:ai:runtime'), 'llama.cpp')
      )
      and $number($nullSafeProp($, 'cdx:ai:contextWindow')) >= 131072
    ]
  location: |
    {
      "bomRef": $."bom-ref"
    }
  message: "Local AI model '{{ name }}' advertises a large context window of {{ $prop($, 'cdx:ai:contextWindow') }} tokens"
  mitigation: "Validate that the target environment has enough memory and latency budget for the configured context window, or choose a smaller context budget."
  evidence: |
    {
      "runtime": $prop($, 'cdx:ai:runtime'),
      "contextWindow": $prop($, 'cdx:ai:contextWindow'),
      "artifactFormat": $prop($, 'cdx:ai:artifactFormat')
    }

- id: AIP-002
  name: "Large local AI model lacks quantization metadata"
  description: "Large locally hosted models without quantization metadata are harder to capacity-plan and can hide unnecessary CPU, RAM, or GPU pressure."
  severity: medium
  category: ai-performance
  dry-run-support: full
  standards:
    eu-ai-act:
      - "Technical documentation and traceability"
      - "Performance and deployment characteristics should be documented"
    nist-ai-rmf:
      - "Measure"
      - "Manage"
  condition: |
    components[
      type = 'machine-learning-model'
      and (
        $prop($, 'cdx:ai:source') = 'local-artifact'
        or $prop($, 'cdx:ai:source') = 'local-config'
      )
      and $number($nullSafeProp($, 'cdx:ai:parameterCount')) >= 7000000000
      and $nullSafeProp($, 'cdx:ai:quantization') = ''
    ]
  location: |
    {
      "bomRef": $."bom-ref"
    }
  message: "Large local AI model '{{ name }}' lacks quantization metadata"
  mitigation: "Record quantization and deployment sizing details for large local models so reviewers can validate resource budgets and runtime fit before rollout."
  evidence: |
    {
      "runtime": $prop($, 'cdx:ai:runtime'),
      "parameterCount": $prop($, 'cdx:ai:parameterCount'),
      "source": $prop($, 'cdx:ai:source')
    }