First-class connectors across AI safety, regulatory frameworks, cloud, identity, MLOps, DevOps, and observability — including 5 alerting/SIEM providers (4 more on the roadmap), 4 ML classifier backends, and Python / TypeScript SDK v0.2.0. Roadmap items are badged.
Safety events (policy violations, SLO breaches, drift alerts, human review requests) route to your alerting and SIEM stack. Five providers ship today; four more are on the roadmap.
HEC collector — safety events indexed with structured JSON. Supports custom source type, index, and field extraction for SIEM correlation rules.
Events API v2 — safety violations, SLO burn-rate alerts, and drift events sent as Datadog events. Tag by severity, agent_id, and workspace for dashboard filtering.
REST API v3 — safety incidents auto-create tickets with severity, violation category, agent_id, and remediation steps. Routes to SAFE project (configurable).
Events API v2 — critical safety events (burn_rate_1h > 14.4×, CRIT violations, circuit breaker triggers) page on-call with full event detail and dedup key.
Block Kit messages with Approve / Reject action buttons for human review requests. Severity-colour-coded, links directly to the review queue in the platform.
Log Analytics Workspace ingestion — safety audit events forwarded via the Data Collector API for Sentinel analytic rules and playbooks.
Table API — safety incidents mapped to ServiceNow incidents with priority, category, and description for standard and custom ITIL workflows.
Alerts API — safety-tier severity mapped to OpsGenie priority (P1–P4), integrating with existing escalation policies and team routing.
Adaptive Cards — safety alerts rendered as rich Teams cards with severity badge, violation summary, and direct link to the human review workflow.
Plug in one or more ML classifiers via the SAFETY_CLASSIFIER environment variable. The ensemble runs after the regex pre-screen; the voting strategy (highest severity, majority vote, all-flag, or weighted) is configurable per deployment.
Recommended default — runs entirely on-premises with no API key. Pull llama-guard3 via Ollama for air-gapped deployments. Sub-100 ms inference on GPU.
Azure-hosted classifier for images, text, and multimodal content. Integrates with Azure PIM for access control. Meets EU AI Act Article 9 human oversight requirements.
Zero-infrastructure classifier with broad category coverage. Recommended for production ensembles alongside Llama Guard (highest-severity strategy).
Google's open-weight safety classifier, deployable via HuggingFace or Vertex AI. Configurable via GOOGLE_API_KEY or local inference for data-sovereignty requirements.
Pre-built compliance plugins covering all major AI regulation regimes. Evaluate, generate signed artifacts, and export audit packages in minutes. Automated regulatory reports (Markdown / PDF) generated via RegulatoryReporter.
Articles 9/12/13/17 — risk management system, event logging, transparency documentation, quality management. Automated conformity assessments and prohibited-use classification.
GOVERN / MAP / MEASURE / MANAGE core functions. AI RMF profile generation, action plan, and MEASURE-ME4 adversarial robustness evidence. Automated report generation.
Model risk management, challenger model requirements, validation evidence, ongoing monitoring artifacts. Automated SR 11-7 report for model inventory and validation sign-off.
Open incidents mapped automatically to Plan of Action and Milestones items. Report generation covers audit and accountability controls, configuration management, and incident response.
ICT risk register, incident response artifacts, TLPT references for Critical-risk models. 12 controls evaluated automatically.
Singapore TRMG 2021, Fairness / Ethics / Accountability / Transparency assessment. Cross-border jurisdiction routing in Federation module respects MAS gates per participant.
Clinical AI safety validation, PHI boundary controls, PII masking via Presidio, SaMD submission artifacts. Routes PHI prompts to on-premises endpoints automatically.
Material AI risk disclosure, control testing, board-ready audit packages. CompliancePipelineAgent auto-promotes or queues for Legal HITL review.
Native IDE integration, typed SDKs for Python and TypeScript, and CI/CD gates that bring AI safety and governance into every step of the development workflow.
31 governance tools across 5 modules — Compliance, Registry, Audit, Monitoring, and AI Safety. Ask Claude: 'validate this prompt through the safety engine', 'show SLO burn rates', 'run the red-team certification suite'. Natural language, no context switching.
aegissovereign — typed sub-clients for all platform surfaces. oss.safety covers the safety subsystems: validate_input, mask, check_model, run_redteam, list_reviews, submit_review, generate_report, slo_summary, and more.
@aegissovereign/sdk — 30 new typed safety types, client.safety sub-client. Works in Node.js 18+ and modern browsers. Covers validateInput, checkModel, runRedteam, getSLOSummary, submitReview, generateReport, and more.
compliance-action@v1 — regulatory eval gates on every PR. safety-redteam.yml — nightly CI job: static certification (18+ probes), multi-turn attack chains, LLM-generated probe mutations, NoveltyTracker regression detection.
Connect your existing ML experiment tracking and training infrastructure to the governance pipeline. Every external model import triggers CompliancePipelineAgent — regulatory evals, bias checks, adversarial robustness, and the AI safety gate run automatically.
Experiment ID linking, training run lineage, model artifact URI — registered automatically on model card generation. Safety gate runs on every registered model.
Run metadata pulled into model registry; bias, robustness, and safety scores synced back to W&B runs for unified reporting.
Import models directly from Hub; model card populated from Hub metadata, compliance eval and safety gate triggered on import. Provenance lineage recorded back to the base model.
Bi-directional sync — compliance scores, bias results, and Model Safety Registry rating surfaced in SageMaker console alongside SageMaker model versions.
Model registration from Azure ML assets; compliance evaluations linked to Azure ML experiment runs via REST API. Safety eval results written to model metadata.
Self-hosted LLM endpoint for air-gapped deployments. Used as the LLM gateway provider for PII / PHI classification routing and as the Llama Guard classifier backend.
One universal Helm chart deploys across every major cloud with native integrations for secrets management, IAM, SIEM, and storage. No preferred cloud — your data stays on your infrastructure.
AKS, Azure AD / Entra ID, Key Vault (secret key provider), Azure Monitor, Azure Sentinel (safety SIEM), Blob Storage, Azure AI Content Safety (ML classifier).
EKS, IAM / IRSA, Secrets Manager (secret key provider), Security Hub, EventBridge, CloudTrail, S3. Bedrock via LLM Gateway provider.
GKE, Cloud IAM / Workload Identity, Secret Manager (secret key provider), Cloud Logging, Security Command Center, Pub/Sub, GCS. Vertex AI / Gemini via LLM Gateway.
Vanilla Kubernetes, MetalLB, HashiCorp Vault (secret key provider), Ollama (local LLMs + Llama Guard). Full Sovereign plan — no outbound internet required.
Enterprise-grade identity providers with OIDC/PKCE and SCIM 2.0 provisioning. Reviewer qualifications for the human review workflow are derived from JWT claims at login.
SSO, MFA, Conditional Access, SCIM 2.0 provisioning. reviewer_qualifications JWT claim auto-populates HITL reviewer clearance levels.
SAML 2.0, SCIM provisioning, adaptive MFA. Okta group memberships map to platform RBAC roles (admin, ops, legal, dev).
Universal Login, RBAC, social connections. JWT claims include trust_level used by the multi-agent trust boundary enforcement.
Pre-configured realm included in the Helm chart — zero external IdP dependency for air-gapped deployments. Realm export/import for DR.
Connect to your existing data infrastructure with certified connectors — including column-level PII lineage tracking and GDPR-native erasure request propagation.
Column-level lineage, warehouse query hooks, PII propagation tracking. GDPR erasure requests propagate to source tables via change streams.
Dataset lineage DAG, PII column tagging, audit export to GCS. Safety violation events can be streamed to BigQuery for long-term analytics.
Primary datastore — row-level security, connection pooling via pgBouncer, WAL archiving via wal-g to S3/GCS for PITR. Audit chain stored with row-level delete prevention.
Session caching, Celery broker (AI agent task queue), rate limiting, pub/sub for real-time safety event stream. Redis Streams used for live violation feed.
Automate deployments and enforce compliance gates in your existing toolchain. The platform ships as a production-ready Helm chart with pre-upgrade migration hooks.
Helm chart with pre-upgrade hook (alembic migrate), liveness/readiness probes, HPA, and NetworkPolicy templates. Supports AKS, EKS, GKE, and vanilla K8s.
Infrastructure-as-Code modules for AKS, EKS, GKE, and on-prem. State management via Azure Blob, S3, or GCS backends.
Universal chart with values overlays per environment. Single command deploy: helm upgrade --install aegissovereign helm/sovereign-gateway --values values-production.yaml
Full-stack monitoring with industry-standard platforms. The platform publishes Prometheus metrics for platform health and safety-guard decisions; deeper safety analytics (violation rates, SLO burn, review queue depth) are computed from real audit events via the safety analytics API.
Pre-built Grafana dashboard (grafana/opensovereign-dashboard.json). Metrics under the opensovereign_* prefix — agent runs, HITL queue depth, Celery queues, CPU/GPU — plus aegis_safety_guard_decisions_total for guard outcomes.
APM, log management, custom dashboards. Safety events routed via SafetyIntegrationHub — separate from the APM stream. Safety SLO burn rates can be mapped to Datadog SLOs.
Distributed tracing, vendor-neutral telemetry, W3C Trace Context propagation across agent graph executions. Safety pipeline trace spans included.
HEC — audit events and safety events indexed with structured JSON. Platform audit chain entries and safety violations in separate source types for differentiated SIEM rules.