LLM guardrails are checks placed around a language model to catch unwanted input or output: toxic language, off-topic requests, leaked personal data, jailbreak attempts. They're an important layer. But as teams move from chatbots to agents that take actions, it's worth being precise about what LLM guardrails cover — and what they don't. This guide is part of our series on AI agent security.
What LLM guardrails do well
Most guardrail tools work on text, before or after the model:
- Input guardrails screen what goes into the model — blocking jailbreak patterns, off-topic requests or abusive content.
- Output guardrails screen what comes out — catching personal data, unsafe advice, competitor mentions, or answers that don't match a schema.
- Topic and tone controls keep an assistant within its intended scope.
For customer-facing chat, these are essential. They reduce embarrassing answers and catch a lot of low-effort misuse.
What LLM guardrails miss once agents act
An agent's riskiest output isn't a sentence — it's a tool call. Text-focused guardrails often don't see:
- The action and its arguments.
issue_refund(amount=4800)may look perfectly polite as text. - Context outside the conversation. How many refunds has this agent issued today? Is this customer flagged? Is this production?
- Who should decide. Content filters can block or allow; they don't route an action to a finance lead for approval.
- A record of decisions. Many guardrail tools log flags, not a reviewable history of allowed and denied actions.
That gap is why prompts aren't limits — and why text filtering alone isn't either.
Action controls: guardrails for what agents do
Action controls evaluate each tool call before it runs. They look at structured data — the action, its arguments, the agent, the user, the environment — and return one of three answers: allow, deny, or require approval. Because they don't read the conversation, they can't be argued with. Rules are written by the people who own the process and versioned like code. See AI agent policy for how teams structure them.
Use both — at different points
| Content guardrails | Action controls | |
|---|---|---|
| Checks | Text in and out of the model | Tool calls before they run |
| Best at | Harmful, off-topic or leaked content | Money, data, messages, systems |
| Outcome | Block, rewrite, flag | Allow, deny, require approval |
| Resists prompt injection | Partly | Yes — rules sit outside the model |
A sensible stack uses content guardrails on the conversation and action controls on the tools. We compare these and other approaches on our why page.
Choosing LLM guardrails
When evaluating guardrail tools, ask:
- Do they see tool calls and their arguments, or only text?
- Can business owners read and approve the rules?
- Can a rule send an action to a person, not just block it?
- Is every decision recorded with the rule version that made it?
- Can you test a rule change against past activity before it goes live?
Where Gaten fits
Gaten is an action control: it sits between agents and their tools and checks each call against your policy. It works alongside the content guardrails you already use. See how it works.
Frequently asked questions
What are LLM guardrails?
Checks around a language model that screen its input and output for unwanted content, such as jailbreaks, personal data or off-topic answers.
Are LLM guardrails enough for AI agents?
Not on their own. They focus on text; agents need controls on the actions they take, checked before those actions run.
Do guardrails replace human review?
No. Guardrails and action controls handle the routine cases automatically, so people only review the actions that genuinely need judgement.

