← Back to Explorer
Guardrails & Safety
/guardrails
Validates inputs and outputs to ensure safety, policy compliance, and defense against prompt injection.
Strata: L6
Axes: Governance
Status: seed
Explainer
Guardrails are deterministic validation layers embedded directly into the orchestration pipeline. They inspect every input, output, and tool invocation against policy-as-code rules before execution proceeds. Without them, agents are vulnerable to prompt injection, unauthorized privilege escalation, and compliance violations. Guardrails should be baked in from day one, not bolted on after deployment.
Anti-Patterns
- Bolting security guardrails onto an application after deployment results in fragmented policies that agents can easily bypass through prompt manipulation.
Mitigations
- Embed deterministic, policy-as-code guardrails directly into the orchestration layer from day one to validate all inputs, outputs, and tool invocations.