LLM guardrail tools — three different jobs under one word
Sort the category before comparing vendors.
AI Tools Hub Editorial · published · updated
Three layers
| Layer | What it stops | Examples |
|---|---|---|
| Input defence | Prompt injection and malicious input | Lakera |
| Output inspection | Leaked PII, unsafe content, format violations | Lakera, in-house |
| Policy enforcement | Who may send what to which model | Credal, Portkey |
Who needs input defence
Anyone feeding externally sourced data to a model. And only after the permission design limits the damage — an inspection layer alone leaves nothing behind the attacks that get through.
Who needs output inspection
Anyone whose output reaches an end user directly. What to check is domain-specific: PII, competitor names, promises about refunds, delivery dates or legal advice. A vendor's default ruleset will not cover your industry's version.
Who needs policy enforcement
Organisations where several teams use several models and the question is which data may reach which provider. One team on one model does not need this layer; application configuration is enough.
Spend in this order
Permission design (free), in-house output checks (engineering time), policy enforcement (if the org is complex), then an input defence product (if you read external data). Most of this layer is work rather than purchase.
FAQ
How far does in-house get you?
Output checks go a long way with regular expressions and deny lists. Input defence needs ongoing pattern research, which is where a product earns its price.