Robo8AI you can put in front of an auditor.
Robo8's design (explainable verdicts, human-in-command response, a tamper-evident audit trail, poisoning-resistant learning, and local-first data residency) maps directly to the controls enterprise and regulated buyers ask for. This is our control mapping; the live model card and AI bill-of-materials are served from the running system.
Control mapping
Honest split between what the product provides and what remains the operator's responsibility (a SOC 2 report, for example, is an organisational audit, not something software alone delivers).
| Control | Requirement | What Robo8 provides | Operator responsibility |
|---|---|---|---|
| SOC 2: Access (CC6) | Least privilege / logical access | RBAC, SSO (OIDC), SCIM provisioning, hashed tokens, TLS | Run the audit; manage IdP & access reviews |
| SOC 2: Monitoring (CC7) | Detect & respond to anomalies | The product itself: detection, correlation, alerting, drift monitoring | Define monitoring SLAs & on-call |
| SOC 2: Audit (CC7.3) | Complete, reviewable record | Append-only audit of every verdict/action/approval/retrain; identity-bound approvals | Forward to immutable storage/SIEM; set retention |
| EU AI Act: Art. 14 | Human oversight of high-risk AI | Human-in-the-loop; destructive actions need authenticated approval; dry-run default | Keep a human in the loop; don't disable safeguards |
| EU AI Act: Art. 12 | Automatic logging / traceability | Tamper-evident audit + Prometheus metrics across the decision lifecycle | Retain & protect log integrity |
| EU AI Act: Art. 13 | Transparency / interpretability | Explainable verdicts (technique, evidence, confidence, rationale) + model card | Communicate AI use to stakeholders |
| NIST AI RMF: Govern | Policy, accountability, docs | This pack: model card, AI-BOM, control mapping, security policies | Adopt an AI governance policy; assign ownership |
| NIST AI RMF: Manage | Manage risk & data integrity | Poisoning-resistant consensus training, drift detection, augment-never-suppress | Review drift; approve retrains |
| Data residency / GDPR | Control where data lives | Local-first: detection, reasoning & storage on your infrastructure; no egress by default | Sign a DPA; configure retention; choose local vs. cloud LLM |
Model card & AI bill-of-materials
Every deployment exposes a live, machine-readable model card and AI-BOM for the threat
classifier (purpose, features, training provenance, limitations, and safeguards) at
GET /governance/model-card and GET /governance.
- Augment-never-suppress: the model can escalate a detection but never silently drop a heuristic-flagged threat.
- Poisoning defence: training data admitted only by multi-analyst consensus; low-trust contributors excluded.
- Drift monitoring: PSI on the threat-score distribution + rolling accuracy, with triggered retraining.
- Explainability & audit: every prediction carries its reasoning and is logged.
Policy advisor & enforcement
Beyond the control mapping above, Robo8 ships a policy-as-code advisor that
evaluates the live deployment against security policies: response posture, authentication,
TLS, data residency, known-exploited-vuln ownership, incident response, identity handling, model
drift and audit, and returns a scored pass / warn / fail with evidence and remediation, each mapped
to SOC 2 / EU AI Act / NIST. It's advisory by default and
glass-box; turn on enforcement (ROBO8_POLICY_ENFORCE) to mark failing enforceable
policies as blocking, with a human still in command. Teams add their own rules via a JSON overlay.
Available in the console's Policy tab, at GET /policy, and via
python -m robo8.policy_cli.