Skip to content

Key capabilities

Opaque provides a secure, policy-enforced environment for confidential AI. Its capabilities protect data and models throughout execution — before a workflow starts, while it runs, and after it completes.

Before execution: Establishing trust

  • Multi-layer attestation: Opaque verifies the integrity of the compute environment before any workflow runs.
    • Azure confidential node pools validate that host servers are trusted.
    • Opaque trusted computing base (O-TCB) nodes verify that binaries inside the nodes match approved reference measurements.
    • cGPU verification ensures GPUs are attested by NVIDIA.
    • The Opaque SDK checks attestation reports automatically and blocks execution if verification fails.
  • Secret management: Credentials and API keys remain encrypted at all times. They are injected securely at runtime and never displayed in plaintext. Access is governed by workspace roles and cannot be escalated by users.
  • Data connectors: Connectors define approved access points to data or APIs. Users can register, share, and revoke connectors across workspaces. Connectors inherit workspace permissions and provide consistent, auditable data access.
  • PII redaction and unredaction: Configurable redaction nodes remove sensitive fields from prompts and documents before they reach external models. Content can be safely unredacted when returned to a trusted environment.

During execution: Confidential processing

  • Confidential GPU execution: Workflows can run on NVIDIA H100 GPUs with hardware-backed isolation for high-performance confidential AI workloads.
  • Guardrails and policy enforcement: Integrated NVIDIA NeMo Guardrails enforce model and workflow boundaries during execution. Access, data, and runtime policies define who can run workflows, what data they can access, and how models behave.
  • Workflow authoring and iteration: The interactive designer allows users to create, test, and revise workflows within defined policies. Workflows can be reverted to draft for review and change tracking.
  • Hosted and external LLMs:
    • Hosted models can run on vLLM in customer-controlled environments.
    • External APIs (OpenAI, Anthropic) are accessed through Opaque with redaction and guardrails applied automatically.
    • Supported models include Llama 3.2 (1B, 3B) and Llama 2 (7B, 13B).
  • RAG data sources: Integrations with Azure AI Search and Neo4j Cypher support policy-compliant retrieval for retrieval-augmented generation (RAG) workflows.

After execution: Verification and audit

  • Audit logging: All major events—such as logins, workspace creation, asset sharing, and workflow launches—are logged automatically. Logs are cryptographically signed to ensure integrity and traceability.
  • Post-execution attestation reports: Each completed workflow produces a signed attestation report linking the runtime environment, attested components, and applied policies. These reports provide verifiable proof of trusted execution.

Putting it all together

Opaque’s key capabilities ensure that:

  • All environments and components are verified before execution.
  • Data, models, and connectors remain confidential and policy-governed during execution.
  • Every action and workload is auditable after execution.

Together, these features enable enterprises to deploy retrieval-based AI workflows and services over sensitive data with verifiable privacy and compliance.