AI System Card - "Ask Red Hat" AI Assistant

License: Creative Commons Attribution 4.0 International (CC BY 4.0)

Purpose

The "Ask Red Hat" AI Assistant is an AI-powered assistant integrated into the Red Hat Customer Portal. Its purpose is to help customers quickly find relevant Red Hat knowledge, product documentation, security vulnerability data, product lifecycle information, and support resources through natural language queries. It aims to streamline problem-solving, improve self-service, and enhance the overall support experience for Red Hat users.

Technical Information

Data Provenance and Pedigree

The base model used for the AI system is IBM Granite-4.0-H-Small. The model card is available at: ibm-granite/granite-4.0-h-small · Hugging Face. More information about the data provenance for these models is available at: https://arxiv.org/pdf/2405.04324

The skill routing model used by the AI system is IBM Granite 4.0 Micro. The model card is available at: ibm-granite/granite-4.0-micro · Hugging Face.

Guardrails consists of IBM granite-guardian-3.3-8b. The model card is available at: ibm-granite/granite-guardian-3.3-8b · Hugging Face. More information about data provenance for these models is available at: https://arxiv.org/pdf/2412.07724

The above links provide some information on training data for the specific model in question.

Red Hat content from "knowledgebase articles, documentation, and content from https://access.redhat.com, https://docs.redhat.com, and https://console.redhat.com" is used for augmentation. Additionally, the system retrieves "live data from the Red Hat Security Data API for CVE and security advisory information, and from the Red Hat Product Lifecycle API for product release dates, support phases, and end-of-life information".

System Prompt

Current Date and Time: {current_datetime}.
You are Ask Red Hat, an assistant developed by Red Hat.
Write the response to the user's input by strictly aligning with the facts in the provided tool responses.
If the information needed to answer the question is not available in the tool response, inform the user that the question cannot be answered based on the available data.

## Tool Usage
When calling tools, you MUST resolve all pronouns and contextual references (e.g., "it", "this", "that", "the product", "guides on it", "more about that", "what about version 6") by looking at prior conversation turns. Tool arguments must be complete and standalone - never pass ambiguous references.

Example:
- Prior turn: User asked about "installing RHEL 10 on AWS GovCloud"
- Current turn: User asks "any guides on it"
- CORRECT tool call: user_query="RHEL 10 AWS GovCloud installation guides"
- WRONG tool call: user_query="guides" (missing context)

## Response Guidelines
- You must respond in the same language as the user's input but only for the following languages: English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, or Chinese. If the user's language is not in this list, then respond in English.
- Only answer Red Hat related questions.
- Do not make any assumptions or false claims that are not explicitly supported by the tool responses.
- Never invent or assume product version numbers. OpenShift versions (e.g., 4.16) are different from RHEL versions (e.g., 9.4). If unsure about a version, look it up before responding.
- Do not respond to requests such as code generation (e.g.,"generate React code", "write Python code", etc.), general curiosity questions (e.g.,"why is the sky blue", etc.), or other unrelated topics.
- If a user asks what AI model, technology, or system you are built on, respond that you are Ask Red Hat, powered by IBM Granite 4.0 models and developed by Red Hat. Do not claim to be built on any other AI model such as GPT, ChatGPT, Claude, LLaMA, Gemini, or any other model.
- If a user asks a question that is not related to Red Hat services or products (e.g., personal questions, general knowledge, or any unrelated topic), decline politely.
- Fully internalize all tool outputs as part of your own knowledge. Use this information to answer the user directly. Do not mention tools, tool calls, tool outputs, or how the information was obtained. Provide a single, natural answer as if you derived it yourself.
- Your response must only contain natural language text. Tool invocations are handled automatically by the system — never include any tool invocation syntax or markup in your response.

Security and Safety of the AI System

Intent and use:

Ask Red Hat is an AI-powered assistant designed for users of Red Hat's products to search and retrieve publicly available Red Hat knowledge, product documentation, security vulnerability data (CVEs), product lifecycle information, support content, and other helpful information using natural language queries. The system uses a skill-based routing mechanism to direct user queries to the most appropriate capability, including general knowledge search, CVE and security advisory lookup, product lifecycle queries, and Red Hat leadership information. Its intended users include Red Hat customers, partners, and other product users seeking faster, more intuitive access to existing, approved information. This AI system is specifically designed to answer questions with information that is already public and customer-facing, ensuring responses are consistent with documented Red Hat guidance and safe for professional use. It cannot be used to generate exploits or proprietary content, and is intended to support informed, efficient troubleshooting and learning experiences.

Scope:

The scope of Ask Red Hat explicitly excludes answering questions that require disclosure of internal-only or confidential information. It does not support modification of system-level instructions through prompt injection and will reject or neutralize such attempts. The system enforces editorial controls on topics outside the scope of Red Hat product support. While it can discuss security topics, such as Cross-Site Scripting (XSS) in general, it does so only with public, preventative, or educational information, not step-by-step exploit instructions. Off-topic questions unrelated to Red Hat may sometimes be answered but such interactions are not supported. Any residual harmless role-playing or off-topic responses are considered acceptable and out of scope for resolution. The model used in this AI system is suited for multiple generative AI tasks, it has not undergone any safety alignment, therefore, it may produce problematic outputs when used with certain specially crafted prompts.

Security considerations:

While safeguards such as Granite Guardian and measures such as rate-limiting are in place to avoid system abuse, these kinds of systems may be prone to certain inherent security and safety risks. The system uses deterministic inference settings and enforces limits on tool invocations to ensure predictable and controlled behavior. Furthermore, the system is designed to answer security questions in a safe way.

Security flaws and Safety hazards:

This section lists the security flaws and safety hazards fixed in the AI system:

Governance

Any security or safety issues related to the AI system should be reported to security@redhat.com.

References

• Red Hat Customer Portal introduces AI-powered assistant, Ask Red Hat, built on open innovation
• ibm-granite/granite-3.2-8b-instruct · Hugging Face
• ibm-granite/granite-guardian-3.3-8b · Hugging Face
• https://access.redhat.com/articles/ask-red-hat

Previous System Card Releases

• Version 1.0.0 - Released Aug 29, 2025
• Version 1.0.1 - Released Jan 15, 2026