HumAi HQ AOS shipped a platform-wide safety layer this week: NVIDIA NeMo Guardrails, running as a service beside the runtime, checking every agent turn against NVIDIA's NemoGuard content-safety model.
What actually runs
Two checks wrap every turn. Before an agent acts, the operator's message passes an input rail. Before a reply is delivered, or absorbed into the agent's long-term memory, it passes an output rail. The safety model is an 8B NemoGuard guard screening 23 harm categories, from violence and malware through to PII leakage, at about a second per check.
The defaults are deliberate. The input rail monitors and audit-logs rather than blocking the person typing, because a console that nannies its operator does not get used. The output rail blocks: an unsafe reply is withheld and replaced with a notice, and the event lands in the same immutable audit trail as everything else on the platform, filterable under Safety Rails.
Admin control, per agent
Every agent's config page now carries a Safety Rails card: a master switch, a mode for each rail, and an opt-in topic-control rail for agents that need to stay in their lane, such as customer-facing deployments. Owner and admin roles only. The security posture of a whole fleet is set agent by agent, visibly, rather than buried in a config file.
Fail-open, on purpose
The engineering decision worth noting: the safety service runs beside the platform, not inside the model path, and it fails open. If the sidecar stalls, checks are skipped, the event is logged, and the fleet keeps working. A safety outage can never become a platform outage.
The full picture, including the live control card, is on the HumAi HQ AOS page.