Content guardrails scan both directions of every call — what's sent to the model and what comes back — against built-in threat packs and rules your own admins write. Applied inline, streaming included, with independent control down to the individual billing node.
Prompt-injection, jailbreak, and PII patterns, ready to turn on — protection from day one, no security hire required.
Any tenant admin can add a custom regex-based rule to cover risks unique to the business — leaked project codenames, confidential terms, anything a pattern can catch.
Rules scan what users send and what the model sends back — stopping leaks going out, not only attacks coming in.
Every rule is Disabled Flag Redact Block — match the response to the risk instead of an all-or-nothing switch.
Scope a rule to one team's node, several nodes, or the whole account — different products, different rules, one account.
Restrict a rule to a single API key or member, so a new rule can be trialled safely before it's rolled out to everyone.
A single toggle suspends every guardrail, account-wide — undo a misfiring rule instantly, with zero downtime.
Live match counters on every rule, built-in and custom, so you can tune false positives with real traffic data, not guesswork.
Redaction works correctly on live streamed responses without corrupting the stream — production-grade for real chat products.
Guardrails are one layer of defense-in-depth, configured per tenant from the admin console — they run alongside, not instead of, your own application-level checks.