AI Runtime Security Tools

Cloud provider built-in runtime controls

Amazon Bedrock Guardrails, Azure AI Content Safety with Prompt Shields and Google Model Armor each screen prompts and responses for harmful content, sensitive data and prompt attacks. None of the three pages read on 16 Nov 2026 describes blocking shell commands on a developer machine.

Facts read on .

What do Azure, AWS and Google Cloud include?

Each provider ships a content and attack filter for model inputs and outputs. AWS documents Bedrock Guardrails with content filters including a Prompt Attack category, denied topics, word filters and sensitive information filters that block or mask personal data. All three are managed services you call over an API, so they add a network hop and a dependency to each request, which is worth measuring before launch.

Azure documents Prompt Shields, which detect user prompt attacks and document attacks. Google documents Model Armor, which screens prompts and responses for prompt injection, jailbreaks, sensitive data and harmful content.

Sources: Amazon Bedrock Guardrails components; Prompt Shields in Azure AI Content Safety; Model Armor overview, Google Cloud.

How do they enforce?

Each can be called by your application. Bedrock Guardrails can be applied during a model call or invoked alone through the ApplyGuardrail API without calling a model.

Model Armor offers an inspect-only mode and an inspect-and-block mode, so you can test before blocking. Microsoft says Prompt Shields in Foundry can block inputs classed as attacks before the model processes them.

Sources: Amazon Bedrock Guardrails; Model Armor overview, Google Cloud; Prompt Shields in Azure AI Content Safety.

Do they work outside their own cloud?

Some do. Google says Model Armor can be used whether you deploy AI in Google Cloud or on other cloud providers. Bedrock's ApplyGuardrail API takes text from any source.

Check the data path: text leaves your application to be screened, so review residency and logging.

Sources: Model Armor overview, Google Cloud; Amazon Bedrock Guardrails.

When is a separate tool needed?

Add a separate tool when the harm is an agent action, not text: a shell command, a file read, a package install or a network call. The three provider pages read on 16 Nov 2026 are about prompts and responses.

Also add one when you need endpoint coverage, a cross-cloud audit trail in one place, or enforcement in agents that run outside the provider.

Sources: Amazon Bedrock Guardrails components; What is Azure AI Content Safety; Model Armor overview, Google Cloud.

How do you combine them?

Use the provider control as the base layer for model calls in that cloud, then add agent and endpoint controls where actions happen. Keep one owner for the combined policy so rules do not conflict.

Microsoft also documents Defender for Endpoint runtime protection for agents, currently public preview, which shows a provider can extend past text.

Sources: What's new in Microsoft Defender for Endpoint.

What do the pages not say?

The pages read on 16 Nov 2026 give no latency figure in milliseconds for these controls on the pages sampled, and the failure behavior for a service outage is something to confirm in the service terms or by testing.

Ask your cloud account team for both before relying on the control in a customer path.

Sources: Amazon Bedrock Guardrails; What is Azure AI Content Safety.

What does a rollout look like?

Create a policy template, run it in log-only mode on a copy of traffic, review hits, then enforce. Google's Model Armor documentation suggests separate templates for input and output, so traceability and tuning stay simple.

Record the policy version with each decision.

Sources: Model Armor overview, Google Cloud.

What should you check in each service?

Check regions, quotas, pricing and how logs are stored. Check whether the filter supports your languages and whether sensitive text is stored by the service.

Review the provider's documentation again each quarter, since these services change often.

Sources: Amazon Bedrock Guardrails; What is Azure AI Content Safety; Model Armor overview, Google Cloud.

Sources