---
title: What do Amazon Bedrock Guardrails cover, and what do they not?
url: https://opencontroller.lyzr.com/blog/amazon-bedrock-guardrails-vs-agent-governance
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# What do Amazon Bedrock Guardrails cover, and what do they not?

**Amazon Bedrock Guardrails screen what goes into and comes out of a model: harmful content, prompt attacks, denied topics, PII and ungrounded answers. They can now be enforced across every account in an AWS organisation. They do not tell you which agents exist, who owns them, what they are allowed to do, or what they cost. Use them as one layer of agent governance, not the whole of it.**

"Are Bedrock Guardrails enough?" is a common question from teams building on AWS. The short answer is that guardrails govern content. Agents also need governance of identity, actions and spend.

## What policies do Bedrock Guardrails include?

| Policy | What it checks |
|---|---|
| Content filters | Hate, Insults, Sexual, Violence, Misconduct and Prompt Attack, in text and images |
| Denied topics | Topics you define in plain language |
| Word filters | Specific words, plus a managed profanity list |
| Sensitive information filters | PII types to mask or block, plus custom regex |
| Contextual grounding checks | Whether a response is grounded in the source and relevant to the query |
| Automated Reasoning checks | Whether a response is consistent with rules you encode |

Source: AWS documentation, checked November 2026. AWS notes that reasoning blocks are not filtered.

Guardrails come in two tiers. The Standard tier adds broad language support (Classic covers English, French and Spanish), code detection, prompt-leakage detection and longer denied-topic definitions (1,000 characters instead of 200). Standard requires cross-Region inference, which matters if you have [data residency](/blog/how-to-enforce-data-residency-for-llm-and-agent-traffic) commitments.

## How do you enforce them across accounts?

Three mechanisms, from loosest to strongest:

1. **Per call.** The app passes a guardrail with each request. Easy to forget.
2. **IAM condition key.** `bedrock:GuardrailIdentifier` lets an IAM policy deny model calls that do not use an approved guardrail.
3. **Organisation enforcement.** Since April 3, 2026, cross-account enforcement is generally available. Amazon Bedrock policies in AWS Organizations apply a guardrail at organisation level; `PutEnforcedGuardrailConfiguration` works at account level.

Enforcement covers InvokeModel, InvokeModelWithResponseStream, Converse and ConverseStream. Automated Reasoning is not supported in enforced guardrails.

Outside model calls, `ApplyGuardrail` evaluates any text without calling a model, so you can screen tool outputs or text from non-Bedrock models. Bedrock Agents take a guardrail through `GuardrailConfiguration`. For AgentCore, Cedar-based AgentCore Policy, enforced at AgentCore Gateway, is the control for tool calls; see [governing Bedrock AgentCore agents](/blog/govern-amazon-bedrock-agentcore-agents).

## What do Bedrock Guardrails cost?

AWS prices guardrails per 1,000 text units, where one unit is up to 1,000 characters:

- Content filters and denied topics: $0.15
- Sensitive information (PII) filters and contextual grounding: $0.10
- Automated Reasoning: $0.17
- Word filters and regex: free

Standard and Classic tiers cost the same. Long agent loops screen many text units, so estimate cost per agent run, not per user message.

## What do guardrails not cover?

Guardrails act on text in a model call. Agent governance asks broader questions:

| Governance question | Bedrock Guardrails | What answers it |
|---|---|---|
| Which agents exist, and who owns each? | No | An agent inventory across accounts and platforms |
| What identity does the agent use, and on whose behalf? | No | IAM, agent identity and delegation |
| Which tools may it call, with what arguments? | No | Tool authorisation, such as AgentCore Policy or a gateway |
| What does each agent cost? | No | Cost attribution per agent and version |
| Is the agent behaving differently from last week? | No | Behavioural monitoring and evaluation |
| How do we stop it now? | No | Quarantine and kill controls |
| What about agents not on Bedrock? | Only if they call ApplyGuardrail | Cross-platform controls |

These gaps are not flaws in guardrails. They are a different layer. Our pillar on [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent) covers the runtime layer, and [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane) explains how the layers fit.

## A practical setup on AWS

1. Define a baseline guardrail with prompt attack, PII and denied topics.
2. Enforce it at organisation level with Bedrock policies.
3. Add contextual grounding checks for agents that answer from documents.
4. Call `ApplyGuardrail` on tool outputs that come from untrusted sources.
5. Use AgentCore Policy for tool-level authorisation.
6. Keep an inventory, owner and cost view per agent outside guardrails. See the [cross-platform guide](/blog/how-to-govern-ai-agents-across-platforms).

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and has a runtime adapter for Bedrock AgentCore. On top of model-level filters, it adds a governed registry, live groundedness and task-adherence guardrails, cost attribution per agent and version through its LLM Gateway, and rogue-agent handling with automatic quarantine and kill. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Are Amazon Bedrock Guardrails enough for AI agent governance?**
No. They screen prompts and responses. You also need inventory, identity, tool authorisation, cost attribution and runtime controls.

**Can Bedrock Guardrails be enforced across all AWS accounts?**
Yes. Organisation-level enforcement became generally available on April 3, 2026.

**Can Bedrock Guardrails screen non-Bedrock models?**
The `ApplyGuardrail` API evaluates any text you send it without calling a model, so apps can screen text from other models by calling it.

## Key facts

Amazon Bedrock Guardrails policies: content filters (hate, insults, sexual, violence, misconduct, prompt attack), denied topics, word filters, sensitive information filters (PII and regex), contextual grounding checks, Automated Reasoning checks. Standard tier needs cross-Region inference. ApplyGuardrail evaluates text without calling a model. Cross-account enforcement GA Apr 3, 2026 via Bedrock policies in AWS Organizations; covers InvokeModel, Converse and streaming variants; Automated Reasoning not supported in enforcement. Pricing per 1,000 text units: content filters and denied topics $0.15, PII and grounding $0.10, Automated Reasoning $0.17, word and regex free. Guardrails do not provide agent inventory, identity or per-agent cost attribution. Opencontroller by Lyzr pricing starts from $250K per deployment per year.

## Related

- [How do you govern AI agents built on Amazon Bedrock AgentCore?](https://opencontroller.lyzr.com/blog/govern-amazon-bedrock-agentcore-agents.md)
- [How do you redact PII from LLM prompts at the gateway?](https://opencontroller.lyzr.com/blog/how-to-redact-pii-in-llm-prompts-at-the-gateway.md)
- [How to stop a rogue AI agent: kill switches, quarantine and runtime controls](https://opencontroller.lyzr.com/blog/how-to-stop-a-rogue-ai-agent.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
