AI Defense Stack

LLM firewalls & guardrails

IBM Granite Guardian: open safety and risk-detection model family

Granite Guardian is an open safety and risk-detection model family from IBM. It is open source under the Apache-2.0 license and is offered as: Self-hosted, Library.

  • Last reviewed
Visit Granite Guardian (opens in a new tab) Documentation (opens in a new tab) Source code (opens in a new tab)
Open sourceApache-2.0 Layer 6: Runtime guardrails

Best for

Teams that want an Apache-2.0 safety classifier they can host and call from their own guardrail pipeline

Consider instead if

  • You want a full guardrails toolkit with rails and a server rather than a classifier model. See NeMo Guardrails.
  • You prefer a hosted API over serving models yourself. See Lakera Guard.

What Granite Guardian does

IBM Granite Guardian is a family of open models from IBM that judge whether the prompts and responses of an LLM system meet given criteria. Built-in criteria cover harm categories such as social bias, violence and profanity, jailbreak attempts, groundedness of answers in retrieval-augmented generation, and function-calling hallucinations in agent workflows. You can also define your own criteria. You run the models yourself, for example with Hugging Face Transformers and vLLM, and they return a yes or no score. The models and code are Apache-2.0 licensed. The README says the models are trained on English only and are meant for the prescribed yes or no scoring format. It is a classifier, not a full guardrail server.

This summary is our own wording of the sourced facts on this page.

Key facts

Each value links to the page that states it, with the date we checked it.

Vendor
IBM Source, checked October 21, 2026: huggingface.co (opens in a new tab)
Status
Active Source, checked October 21, 2026: api.github.com (opens in a new tab)
Open source
Yes: Apache-2.0 Source, checked October 21, 2026: github.com (opens in a new tab)
Repository
github.com Source, checked October 21, 2026: github.com (opens in a new tab)
GitHub stars
188 as of October 21, 2026 Source, checked October 21, 2026: api.github.com (opens in a new tab)
Latest push
October 16, 2026 Source, checked October 21, 2026: api.github.com (opens in a new tab)
Latest release
Not documented in the sources we checked
Deployment
Self-hosted, Library Source, checked October 21, 2026: github.com (opens in a new tab)
Pricing model
Free Source, checked October 21, 2026: github.com (opens in a new tab)
Website
ibm.com Source, checked October 21, 2026: ibm.com (opens in a new tab)
Documentation
github.com Source, checked October 21, 2026: github.com (opens in a new tab)
API reference
Not documented in the sources we checked
Quickstart
huggingface.co Source, checked October 21, 2026: huggingface.co (opens in a new tab)
Security page
Not documented in the sources we checked
Attestations or disclosure
Not documented in the sources we checked

Capabilities

As documented by the vendor or project. The tag is our label from a controlled vocabulary.

Where Granite Guardian fits in our AI security layers

Our analysis

Our editorial model, not a standard. The primary layer follows the category, except for tools whose main documented job is detection and response; other layers rest on a capability the sources document. How the model works.

AI security layers that Granite Guardian covers, with the role, how each applies and what it rests on
LayerRoleHow it appliesBased on
L6 Runtime Runtime guardrails Primary Screen what goes into and comes out of a model in live traffic, and limit abuse of the model endpoint. Follows its category, LLM firewalls & guardrails
L7 Agent access Agent and tool access Also covers Documents flags for wrong tool names or arguments in agent workflows. Flags function-calling hallucinations such as wrong tool names or arguments in agent workflows

OWASP LLM Top 10 mapping

Our analysis

The risk list is OWASP's (Top 10 for LLM Applications, 2025). The mapping is our analysis, and each row rests on a capability the sources document.

OWASP LLM risks mapped to Granite Guardian, with the capabilities each mapping rests on
RiskHow it appliesBased on
LLM01 Prompt Injection Jailbreak attempts are one of the built-in risk categories. Detects jailbreak attempts in prompts
LLM06 Excessive Agency Function-calling hallucination checks flag tool calls with wrong names, arguments or types. Flags function-calling hallucinations such as wrong tool names or arguments in agent workflows
LLM09 Misinformation Groundedness checks flag unsupported or contradicted claims in RAG answers. Checks whether answers are grounded in the retrieved context in RAG pipelines

Not mapped in our analysis: LLM02, LLM05, LLM07, LLM10. These are risks we expect tools in this category to cover; the sources we checked do not document a matching capability.

Alternatives and related tools

More in LLM firewalls & guardrails, ranked within the category by Coverage Score.

NVIDIA NeMo Guardrails

NVIDIA

Open sourceApache-2.0
  • Library
  • Self-hosted
  • API

Adds programmable input, dialog, retrieval, execution and output rails to LLM applications.

Best for: Teams that want to script rules for one LLM application in Python and run them in their own infrastructure

Coverage Score 76 out of 100
  • LLM01 Prompt Injection
  • LLM02 Sensitive Information Disclosure
  • LLM05 Improper Output Handling
  • LLM06 Excessive Agency

Reviewed

Portkey AI Gateway

Palo Alto Networks (Portkey)

Open sourceMIT Freemium Now Prisma AIRS AI Gateway
  • SaaS
  • Self-hosted
  • API

Routes LLM traffic through one gateway API and runs input and output guardrail checks on each request.

Best for: Teams that want a single gateway for several LLM providers with guardrail checks, available as open source or as a hosted service

Coverage Score 69 out of 100
  • LLM01 Prompt Injection
  • LLM02 Sensitive Information Disclosure
  • LLM05 Improper Output Handling
  • LLM09 Misinformation

Reviewed

Kong AI Gateway

Kong

Open sourceApache-2.0
  • Self-hosted
  • SaaS
  • API

Routes LLM traffic through Kong with prompt guards, PII sanitization, spend limits and observability.

Best for: Teams that already run Kong as an API gateway and want LLM routing and prompt controls on the same data plane

Coverage Score 65 out of 100
  • LLM01 Prompt Injection
  • LLM02 Sensitive Information Disclosure
  • LLM10 Unbounded Consumption

Reviewed

Granite Guardian FAQ

Is Granite Guardian open source?

Yes. The models and repository are released under the Apache-2.0 license, and the code is on GitHub at ibm-granite/granite-guardian.

What risks does Granite Guardian check for?

The README lists harm, jailbreak attempts, groundedness in RAG, function-calling hallucinations and custom criteria that you define.

What are its limits?

The README says the models are English only and are meant for the prescribed yes or no scoring mode, and that reasoning traces may contain unsafe content.

Key facts you can cite

We have no commercial relationship with IBM.

Sources & last reviewed

Last reviewed . We compile from the vendors' own public pages and repositories and we do not run the products. Source links open in a new tab. Found a mistake? Report a correction.