| name | mandate-2-2-5-runtime-evidence-intake-and-guidance |
| description | Collect natural-language context and generate supporting evidence guidance for Mandate 2.2.5 Model Asset Protection. Use when evaluating operational controls such as encryption, key management, access governance, and exfiltration protections. |
Mandate 2.2.5 Runtime Evidence Intake and Guidance Skill
Mandate
- ID: 2.2.5
- Title: Model Asset Protection
Mitigates
- ML05 Model Theft
- ML10 Model Poisoning
- ML03 Model Inversion Attack
- LLM10 Unbounded Consumption
Objective
Transform user-provided operational context into an actionable evidence collection plan for model confidentiality, integrity, and access governance controls.
Natural-Language Intake Workflow
- Ask where model artifacts are stored, served, and backed up.
- Ask which users/services can read, write, export, or deploy model assets.
- Ask how encryption at rest/in transit is implemented and audited.
- Ask how keys are generated, rotated, and access-controlled.
- Ask how data/model exfiltration is detected and blocked.
- Ask how model integrity/signing is verified in delivery pipelines.
- Ask for incident response process for theft/tampering events.
Context Normalization Schema
Capture and normalize answers into:
asset_inventory: registry locations, buckets, artifact types, endpoints.
access_surface: principals, roles, permissions, boundary controls.
crypto_controls: encryption modes, TLS posture, key ownership.
key_management: KMS/HSM usage, rotation cadence, key policy.
exfiltration_controls: egress restrictions, anomaly detection, DLP hooks.
integrity_chain: signing, checksum validation, attestation in pipeline.
ir_readiness: runbooks, drills, escalation, forensic logging.
Supporting Documents to Request
- Model registry/storage security configurations.
- KMS/HSM policy and key rotation evidence.
- Access and egress audit logs.
- Deployment pipeline attestation/signing records.
- Incident response runbooks and exercise evidence.
Guidance to Generate Supporting Evidence
- Generate model asset inventory.
- Export registry metadata and storage object inventories.
- Generate encryption and key-management evidence.
- Export bucket/volume encryption configs and key policy details.
- Export key rotation status and access logs.
- Generate access governance evidence.
- Export IAM/RBAC bindings and recent read/export activity.
- Generate exfiltration monitoring evidence.
- Export egress logs, anomaly alerts, and blocked events.
- Generate integrity evidence.
- Export signature verification artifacts and pipeline attestations.
- Generate incident readiness evidence.
- Export runbooks, tabletop results, and post-incident reviews.
System Command Templates (Adapt to Environment)
- AWS model storage encryption:
aws s3api get-bucket-encryption --bucket <model-bucket> > bucket_encryption.json
- AWS KMS key posture:
aws kms describe-key --key-id <key-id> > kms_key_details.json
aws kms get-key-rotation-status --key-id <key-id> > kms_rotation_status.json
- IAM and access policy export:
aws iam list-roles > iam_roles.json
gcloud projects get-iam-policy <project-id> --format=json > gcp_iam_policy.json
az role assignment list --all --output json > azure_role_assignments.json
- Access activity logs (example):
aws cloudtrail lookup-events --start-time <start> --end-time <end> > cloudtrail_model_access.json
- Container/image signing checks:
cosign verify --key <public-key> <image-ref>
cosign verify-blob --key <public-key> --signature <sig-file> <artifact-file>
- Kubernetes egress/network controls:
kubectl get networkpolicy -A -o yaml > network_policy_export.yaml
Evidence Completeness Rules
- Require direct evidence for confidentiality, integrity, and access controls.
- Reject evidence without timestamps, owner/service identity, or environment scope.
- Flag evidence as stale if outside the assessment period.
Final Assessment Readiness
- Ready only when storage, crypto, key, access, egress, integrity, and incident-readiness evidence are all present and consistent.
- Not ready if any control domain is missing or weakly evidenced.
Output Contract
Return:
context_profile
asset_control_matrix
required_artifacts_checklist
artifact_generation_steps
command_templates
evidence_status_matrix
assessment_readiness
remaining_gaps
Guardrails
- Keep command guidance read-only unless user explicitly asks for changes.
- Ask follow-up questions for unclear ownership or environment boundaries.
- Avoid final pass/fail until evidence completeness reaches ready state.