Skip to content

feat: model serving + GPU cluster security skills#7

Merged
msaad00 merged 2 commits intomainfrom
feat/model-serving-gpu-security-skills
Apr 9, 2026
Merged

feat: model serving + GPU cluster security skills#7
msaad00 merged 2 commits intomainfrom
feat/model-serving-gpu-security-skills

Conversation

@msaad00
Copy link
Copy Markdown
Owner

@msaad00 msaad00 commented Apr 9, 2026

Summary

Two new security benchmark skills filling critical gaps in AI infrastructure security — no CIS benchmark exists for either domain today.

model-serving-security (16 checks, 31 tests)

Audits model deployment infrastructure: API gateways, Kubernetes serving pods, cloud-native endpoints.

Domain Checks Key Controls
Auth & RBAC 3 Endpoint auth required, no hardcoded secrets, role-based access
Abuse Prevention 2 Rate limiting, input size/token limits
Data Egress 3 Output filtering, memorization guard, PII redaction in logs
Container Runtime 3 No privileged, read-only rootfs, non-root user
TLS & Network 2 TLS enforced, no public endpoints
Safety Layers 3 Prompt injection guard, content safety, model version pinning

Mapped to: MITRE ATLAS (7 techniques), NIST CSF 2.0, OWASP LLM Top 10

gpu-cluster-security (13 checks, 31 tests)

Audits GPU compute clusters: Kubernetes, Docker, bare-metal.

Domain Checks Key Controls
Runtime Isolation 3 No privileged GPU pods, device plugin (not /dev mounts), no host IPC
Driver & CUDA 2 Known CVE driver check (6 versions), CUDA >= 12.2
Network 2 InfiniBand tenant segmentation, NetworkPolicy on GPU namespaces
Storage 2 /dev/shm size limits, model weight encryption at rest
Tenant Isolation 2 Namespace per tenant, GPU resource quotas
Observability 2 DCGM monitoring, GPU audit logging

Mapped to: MITRE ATT&CK (7 techniques), NIST CSF 2.0, CIS Controls v8

Both skills include

  • SKILL.md with Anthropic-spec frontmatter + Mermaid diagrams
  • Security guardrails (read-only, no API calls, safe for production)
  • Human-in-the-loop policy: automated assessment, human required for remediation
  • JSON/console output with CI-friendly exit codes
  • Full test suites (62 tests total)

Also includes

  • CI fix: PYTHONPATH + testpaths override for skill-level pytest
  • CI jobs for both new test suites
  • README + CLAUDE.md updated

Test plan

  • model-serving-security: 31 tests passing
  • gpu-cluster-security: 31 tests passing
  • ruff check + format passing
  • CI workflow includes both new test jobs

msaad00 added 2 commits April 8, 2026 21:04
Two new skills filling gaps in AI infrastructure security:

model-serving-security (16 checks, 31 tests):
  Auth & RBAC (3): endpoint auth, hardcoded secrets, role-based access
  Abuse prevention (2): rate limiting, input size limits
  Data egress (3): output filtering, memorization guard, PII in logs
  Runtime (3): no privileged, read-only rootfs, non-root user
  Network (2): TLS enforced, no public endpoints
  Safety (3): prompt injection guard, content safety, model versioning
  Mapped to: MITRE ATLAS, NIST CSF 2.0, OWASP LLM Top 10

gpu-cluster-security (13 checks, 31 tests):
  Runtime isolation (3): no privileged GPU, device plugin, no host IPC
  Driver/CUDA (2): known CVE check, CUDA version compliance
  Network (2): InfiniBand segmentation, NetworkPolicy on GPU namespaces
  Storage (2): /dev/shm limits, model weight encryption
  Tenant (2): namespace isolation, GPU resource quotas
  Observability (2): DCGM monitoring, audit logging
  Mapped to: MITRE ATT&CK, NIST CSF 2.0, CIS Controls v8

Both skills include:
- Mermaid architecture diagrams
- Security guardrails (read-only, no API calls, safe for production)
- Human-in-the-loop policy (automated assessment, human for remediation)
- Compliance framework mappings with specific control IDs
- JSON/console output with exit codes for CI/CD
- Full test suites (62 tests total, all passing)

CI updated to run both new test suites.
README + CLAUDE.md updated with new skills.
@msaad00 msaad00 merged commit 64551fb into main Apr 9, 2026
7 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant