Learn
What are AI guardrails?
AI guardrails are runtime policies that constrain what a model or agent can output or do — blocking unsafe content, PII leaks, prompt injections, and out-of-scope tool calls. On Neuro OS they include input filters, output checks, and action-level Ask gates before external writes.
In depth
Alignment training makes frontier models helpful by default, but production needs an independent safety layer. Guardrails assume the model will occasionally fail and stop harm before it reaches users or systems.
Four layers: (1) Input — block injections, off-topic queries, sensitive user data before the model runs. (2) Output — scan responses for policy violations, PII leaks, schema errors. (3) Action — the critical layer for agents: allow-lists of tools, rate limits, cost caps, and human approval before send/post/pay. (4) Conversational — topic scope and escalation to humans.
Production stacks combine fast filters on the hot path with slower policy checks for high-stakes actions. Defense in depth — multiple independent checks with different failure modes — beats any single classifier.
On Neuro OS, guardrails are a product surface: Governance ships Ask as default-deny for external actions, scoped connectors, and budgets — not an invisible filter you hope works.
Examples
- Ask gate — agent drafts customer email; human approves before send
- Tool allow-list — role may read CRM but not delete records
- PII redaction — strip emails and card numbers from run logs
- JSON schema validation — structured agent output must match expected shape
- Retrieved-chunk scan — RAG documents checked for injection payloads
- Budget cap — role stops when inference spend hits weekly limit
Related terms
FAQ
Don't aligned models already refuse unsafe requests?
Mostly — but jailbreaks, prompt injections from retrieved docs, and accidental PII leaks still happen. Guardrails enforce your policies even when the model is compromised.
What's the difference between alignment and guardrails?
Alignment is training-time behavior. Guardrails are runtime code — narrow, composable, and auditable. You need both for production agents.
Do guardrails hurt latency?
Fast filters add tens of milliseconds. LLM policy checks add seconds and belong on high-stakes actions only. Clear 'needs approval' messages beat silent refusals.
Run governed agent roles on a company OS — not only definitions in a glossary.