Get in Touch

Course Outline

Offline EXO Deployment

  • Leveraging EXO_OFFLINE to block runtime internet connectivity.
  • Pre-loading models into EXO_MODELS_READ_ONLY_DIRS from trusted internal mirrors.
  • Validating model weight integrity via SHA-256 checksums and signed model cards.
  • Operating EXO within air-gapped networks, independent of HuggingFace services.

Dashboard and API Access Control

  • Setting up and configuring reverse proxies (such as nginx or Caddy) with TLS termination.
  • Deploying role-based access control mechanisms for the EXO dashboard and REST API.
  • Utilizing macOS Keychain or Linux pass to securely store credentials for API authentication.
  • Limiting access to administrative endpoints to designated source IP ranges.

Cluster Isolation and Network Security

  • Segmenting EXO clusters using EXO_LIBP2P_NAMESPACE and VLAN configurations.
  • Configuring host firewalls (including macOS Application Firewall, iptables, and nftables) for EXO ports.
  • Mitigating risks associated with unauthorized device discovery and rogue node injection.
  • Encrypting libp2p traffic between nodes in environments where RDMA is unavailable.

Model Governance and Provenance

  • Establishing an internal model registry that includes approved model lists and associated metadata.
  • Applying tags and versioning to quantized weights (4-bit, 8-bit) in tandem with source checkpoints.
  • Restricting loading capabilities to specific HuggingFace repositories or internal artifacts only.
  • Maintaining documentation on model lineage, licensing terms, and acceptable use policies.

Audit Logging and Compliance

  • Directing EXO logs to immutable audit trails, such as SIEM systems or WORM storage.
  • Correlating API call logs with user identities and precise timestamps.
  • Recording key events, including model instance creation, deletion, and inference requests.
  • Generating periodic compliance reports suitable for internal and external auditors.

Threat Modeling and Incident Response

  • Identifying potential threats such as data exfiltration via model outputs, prompt injection, and side-channel leaks.
  • Implementing pipelines for prompt monitoring and content filtering.
  • Developing incident response runbooks tailored for cluster compromise scenarios.
  • Isolating affected nodes, preserving forensic data, and reconstructing clean environments.

Physical Security and Hardware Boundaries

  • Protecting Thunderbolt ports against unauthorized RDMA cable connections.
  • Utilizing secure enclaves and Apple Silicon hardware attestation where applicable.
  • Managing physical access to clustered Macs and shared storage resources.
  • Documenting hardware lifecycle management and decommissioning procedures.

Regulatory Considerations

  • Aligning EXO deployments with GDPR, HIPAA, and SOC 2 standards.
  • Ensuring data residency by maintaining inference processes on-premise.
  • Documenting vendor supply-chain risks involving MLX, EXO, and model weights.
  • Preparing for AI governance frameworks, including EU AI Act Article 53.

Requirements

  • Practical experience with EXO or alternative local LLM runtimes
  • A solid grasp of Unix filesystem permissions and network ACLs
  • Knowledge of TLS/SSL certificate management and fundamental encryption concepts

Intended Audience

  • Security engineers
  • Compliance officers
  • AI infrastructure administrators responsible for managing sensitive data
 14 Hours

Number of participants


Price per participant

Testimonials (1)

Upcoming Courses

Related Categories