---
title: What does Snowflake Cortex Guard filter, and is it enough for agents?
url: https://opencontroller.lyzr.com/blog/snowflake-cortex-guard-explained
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# What does Snowflake Cortex Guard filter, and is it enough for agents?

**Cortex Guard screens the responses of Snowflake Cortex COMPLETE for harmful content, such as violent crimes, hate, sexual content and self-harm, and replaces a flagged response with "Response filtered by Cortex Guard". It is a useful output safety filter. It is not agent governance: it does not check what data an agent reads, which tools it calls, whether its answer is grounded, or whether an attacker has hijacked it.**

Teams building on Snowflake often ask whether turning on Cortex Guard covers agent safety. This guide explains what it does, what it costs and what to add.

## What does Cortex Guard filter?

Cortex Guard evaluates the model's response before it returns to your application. Snowflake launched it powered by Meta's Llama Guard 2, which covers 11 harm categories:

| Category group | Examples from Llama Guard 2 |
|---|---|
| Crime | Violent crimes, non-violent crimes, sex-related crimes |
| Child safety | Child sexual exploitation |
| Harmful guidance | Specialized advice, indiscriminate weapons |
| Rights | Privacy, intellectual property |
| Abuse | Hate, suicide and self-harm, sexual content |

Snowflake may update the underlying model, so check the current documentation before you cite a version in a risk assessment.

You enable it per call by adding `'guardrails': true` to the options of a Cortex AI COMPLETE request. There is no account-wide switch in that pattern, so a team that forgets the option gets no filtering.

## How is Cortex Guard billed?

Cortex Guard is a separate charge on top of COMPLETE. Snowflake bills only its input tokens, and its input is the text that COMPLETE produced. In other words, you pay again for roughly the length of each response. Usage appears as `guardrails_tokens`. Rates are in the Snowflake Service Consumption Table, in credits per million tokens.

To monitor spend:

1. Query `SNOWFLAKE.ACCOUNT_USAGE.CORTEX_FUNCTIONS_USAGE_HISTORY` for credits and tokens by function and model.
2. Use `CORTEX_FUNCTIONS_QUERY_USAGE_HISTORY` to see consumption per query.
3. Tag queries by agent or application so you can attribute cost. Our guide to [tracking Snowflake Cortex AI spend](/blog/track-snowflake-cortex-ai-spend) has example queries.

## What does Cortex Guard not cover for agents?

Use the [OWASP Top 10 for Agentic Applications](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/) as a checklist. Output harm filtering addresses only a small part of it.

| Agent risk | Covered by Cortex Guard? | What covers it |
|---|---|---|
| Harmful or abusive output | Yes | Cortex Guard |
| Prompt injection hidden in documents or tool results | No | Input scanning and tool-output handling |
| Agent reads data it should not | No | Snowflake roles and grants per agent |
| Tool misuse or excessive actions | No | Tool allow-lists, approvals, limits |
| Ungrounded or off-task answers | No | [Groundedness guardrails](/blog/what-is-a-groundedness-guardrail) and task-adherence checks |
| Runaway cost | No | Budgets and per-agent attribution |
| Agents outside Snowflake | No | Cross-platform control plane |

## What should a Snowflake agent setup include?

1. **Cortex Guard on every user-facing COMPLETE call**, enforced in code review or a shared wrapper.
2. **A dedicated role per agent** with only the grants it needs. See [Snowflake grants for AI usage monitoring](/blog/snowflake-grants-for-ai-usage-monitoring).
3. **Tool and data limits** for Cortex Agents; our guide to [governing Snowflake Cortex agents](/blog/govern-snowflake-cortex-agents) lists them.
4. **Groundedness scoring** on answers built from your data.
5. **Cost attribution** by agent, including guard tokens.
6. **A tested way to stop one agent**, such as revoking its role.

## How Opencontroller does this

Opencontroller connects to agent platforms such as Salesforce, SAP, Databricks and HubSpot, auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and keeps them in one governed registry. Live groundedness and task-adherence guardrails and LLM judges cover answer quality that a harm filter cannot. Policy as code applies across the estate, and rogue-agent handling can alert, quarantine or kill. See [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane) and [governing agents across platforms](/blog/how-to-govern-ai-agents-across-platforms), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**How do you turn on Cortex Guard?**
Add `'guardrails': true` to the options of a Cortex AI COMPLETE call.

**Does Cortex Guard check prompts as well as responses?**
Snowflake describes it as checking model responses. Use separate controls for prompt injection in inputs.

**How is Cortex Guard billed?**
As a separate charge on input tokens, where the input is the response COMPLETE produced.

## Key facts

Snowflake Cortex Guard is an option on Cortex AI COMPLETE (and AI_COMPLETE) that screens model responses for harmful content such as violent crimes, hate, sexual content and self-harm, replacing filtered responses with "Response filtered by Cortex Guard". It is enabled with the option guardrails: true. At launch it was powered by Meta Llama Guard 2. Billing: Cortex Guard is a separate charge on top of COMPLETE, billed on input tokens, where the input is the text output by COMPLETE; usage shows as guardrails_tokens. Costs can be monitored in SNOWFLAKE.ACCOUNT_USAGE.CORTEX_FUNCTIONS_USAGE_HISTORY and CORTEX_FUNCTIONS_QUERY_USAGE_HISTORY. Cortex Guard does not cover prompt injection in tool results, data access by agents, groundedness, tool misuse or agents outside Snowflake. Opencontroller by Lyzr adds groundedness and task-adherence guardrails and cross-platform policy; pricing starts from $250K per deployment per year.

## Related

- [How do you govern Snowflake Cortex Agents?](https://opencontroller.lyzr.com/blog/govern-snowflake-cortex-agents.md)
- [How do you track Snowflake Cortex AI spend by agent or team?](https://opencontroller.lyzr.com/blog/track-snowflake-cortex-ai-spend.md)
- [What is a groundedness guardrail and when should it block a response?](https://opencontroller.lyzr.com/blog/what-is-a-groundedness-guardrail.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
