Obiguard sits in front of your AI calls as a lightweight proxy. Prompts and responses pass through the inspection pipeline — where policy is evaluated, violations are surfaced, and evidence is written to the ledger. Streamed responses are checked once delivered.
Set OPENAI_BASE_URL (or the equivalent for Anthropic, Bedrock, Vertex) to your Obiguard gateway endpoint. All existing code continues to work.
Each request and non-streamed response passes through the inspection pipeline: PII detection, injection classifiers, and your custom Policy Sets — all evaluated before the call completes. Streamed responses are evaluated after delivery.
If a violation matches a Policy Set, Obiguard takes the configured action — block, redact, flag, or hold for human review. Held requests wait in the Review Queue for approval or rejection.
A lightweight HTTP proxy that speaks OpenAI, Anthropic, AWS Bedrock, Google Vertex, and Azure OpenAI. Evaluates each request and response against your active Policy Sets before forwarding.
Policy Sets group enforcement rules (Criteria) and are assigned to specific AI Agents. Each rule defines a guard or criterion, a scope (prompt, response, or both), and an action. Policy Set versions move from draft through pending review and approved to published.
Every policy breach creates a Violation record with the full event context. Violations can be auto-resolved or routed to the Review Queue for human annotation. All decisions are written to the Audit Log.
A project-scoped registry of every AI Agent and Use Case your team has defined. Agents are linked to Policy Sets — ensuring enforcement follows the workload. Use Cases capture the business intent behind each agent.
Timestamped records of requests, policy decisions, and configuration changes, exportable to CSV or JSON for auditors.
Org-level Controls define what behaviour is prohibited or required. Each Control maps to one or more Criteria, and carries an approval workflow with segregation of duties — a submitter cannot approve their own control.
An embedded Project Moonshot backend from AI Verify Foundation. Benchmark runs (cookbooks and recipes) and red-teaming attack modules execute against a registered agent; api-service resolves the provider connector and injects the token per run, so the portal never handles a credential.
A scheduled job runs at 03:00 UTC, samples ~17% of the previous day’s traces, and scores each against six LLM judge criteria. Fully asynchronous — it reads traces after the fact and never sits on the request path.
A coverage endpoint computes, per framework category, whether your live platform activity evidences it — covered, partial, or not covered — across NIST AI RMF, ISO/IEC 42001, OWASP LLM Top 10, Malaysia’s AISCF, and Singapore’s MGF for GenAI. Each mapping records the standard edition it was made against.
OpenAI, Anthropic, AWS Bedrock, Google Vertex AI, Azure OpenAI, and any self-hosted model behind an OpenAI-compatible API. No special SDK required — just change the base URL.
Export request records, policy decisions, and the Audit Log to CSV or JSON, plus framework coverage as CSV, for an ISO/IEC 42001 audit without manual evidence collection.
Connect in minutes. Content is processed in-memory and never stored. Only policy decisions and metadata are persisted. Ideal for teams that want governance without infrastructure overhead.
Run Obiguard inside your AWS, GCP, or Azure VPC. Outbound traffic is limited to the model providers and services you configure. Enterprise customers receive private container images with a Docker Compose quickstart — your team owns the deployment, we own the updates.
For regulated industries that limit outbound connectivity. Obiguard runs behind an internal image mirror with outbound traffic limited to the destinations you allow. Fully air-gapped operation is on the roadmap.