Guardrails Agent Observability | Splunk

Splunk Agent Observability

Guardrails

Runtime AI guardrails

Stop unsafe agent actions before tools run, using guardrails built from the evaluations you trust.

Block unsafe actions in real time

Catch risky actions the instant they're proposed, and block or redirect them before the tool fires.

Enforce policies before execution, not after

Guardrails sit between the agent and any tool it calls. Each proposed action is checked against your policies, then allowed or blocked in real time. Block prompt injection, PII, PHI, and PCI leakage, tool misuse, and off-path actions before they execute.

Turn evaluations into guardrails

When an agent proposes a tool call, the guardrail evaluates it against your policy and blocks any off-path or unsafe action before anything happens downstream. The same evaluations you already trust become rules enforced at runtime, so there is no second model of correctness to maintain.

Map guardrails to security frameworks

Guardrail controls map to the OWASP Agentic Top 10, so governance teams can see where coverage is established and where it expands, and align enforcement with their own risk framework.

Give compliance teams a model-risk record

Automatically log every guardrail decision: what was proposed, which policy applied, and whether the action was allowed or blocked. That gives model-risk and compliance teams an audit trail they can easily review.

Deploy guardrails in your VPC or fully air-gapped

Deploy guardrails in your own VPC or run fully air-gapped where no data leaves your environment. Keep enforcement and logs under your control to meet residency and isolation requirements.

Watch .conf26 keynotes live

Gain insights from top Splunk minds and industry leaders, September 14–15.

Features

What you can block, and how it is controlled

Explore the documentation

Stop PII, PHI, and PCI before it leaves

Catch sensitive data in a prompt, a response, or a tool call, and block it in real time. Set the rules once and apply them across every agent, so customer records, health data, and card data do not slip out.

Refuse the injected instruction

Detect prompt injection in untrusted input and stop the agent from acting on it. It does not follow a hidden instruction to exfiltrate data or call a tool it should not, because the action is checked before it runs.

Keep agents on the approved path

Block tool misuse and off-path actions. Allow intended actions to proceed while denying risky ones, such as account changes without consent or payment actions over the policy limit.

One control plane for every agent

Get a centralized, framework-agnostic control plane with deny-wins semantics, so the most restrictive decision blocks. It is open source, works with any agent framework, and runs on Luna. Every decision is logged for audit.

AI guardrails FAQs

What are AI agent guardrails?

Guardrails are runtime controls that sit between the agent and the tool. Each proposed action is checked against your policy and then steered or blocked in real time, before anything happens downstream.

How is a guardrail different from monitoring?

Monitoring tells you a bad action happened after the fact. A guardrail acts at the moment of decision, so the risky action is stopped or corrected before the tool runs.

What can guardrails block?

Guardrails block issues like prompt injection, PII, PHI, and PCI leakage, tool misuse, and off-path actions. A fixable action can also be steered; for example, regenerating a response without PII, rather than blocked outright.

Do guardrails work across different agent frameworks?

Yes. A centralized, framework-agnostic control plane with deny-wins semantics enforces the same policy across every agent, and every decision is logged for audit.

How do guardrails support compliance and audit?

Guardrail controls map to the OWASP Agentic Top 10 framework. Every decision is logged as a model-risk record you can hand over, and enforcement can run in your own VPC or fully air-gapped.