AILM-01

Do you limit your solution's LLM privileges by default?

Explanation

This question is asking whether your LLM solution follows the principle of least privilege by default, meaning the AI system only has access to the minimum resources, data, and capabilities necessary to perform its intended functions. In security, the principle of least privilege is fundamental - it limits potential damage if a system is compromised. For LLMs specifically, this means restricting what the model can access, what actions it can perform, and what systems it can interact with. The question aims to understand if your LLM implementation has built-in restrictions that prevent it from: 1. Accessing sensitive data it doesn't need 2. Executing unauthorized commands or code 3. Connecting to systems or services without explicit permission 4. Performing actions beyond its intended use case When answering, you should describe the specific technical controls implemented to restrict the LLM's privileges, such as: - API access limitations - Execution environment restrictions - Data access controls - Authentication and authorization mechanisms - Network isolation measures You should also mention whether these restrictions are enabled by default (without requiring additional configuration) and if they can be modified by administrators.

Guidance

Looking for the LLM tool's privileges and permissions with consideration of the principle of least privilege.

Example Responses

Example Response 1

Yes, our LLM solution implements the principle of least privilege by default The model operates in a sandboxed environment with no direct access to file systems, databases, or external networks unless explicitly granted through our permission framework By default, the LLM can only process the text provided in user prompts and generate responses based on its training data Any additional capabilities such as web browsing, code execution, or integration with other systems require explicit configuration and authentication We implement role-based access controls (RBAC) that govern what actions the LLM can perform, and these permissions are audited regularly Our architecture includes a dedicated orchestration layer that validates all LLM requests against permission policies before execution, ensuring the model cannot exceed its authorized boundaries.

Example Response 2

Yes, our LLM solution employs strict privilege limitations by default The model runs in a containerized environment with read-only access to its own parameters and no network connectivity All interactions with the LLM occur through our API gateway, which enforces authentication, rate limiting, and input validation By default, the LLM has no ability to persist data, execute code, or access external resources When deployed in customer environments, the LLM operates with a dedicated service account that has minimal permissions scoped only to the resources required for its specific use case We provide a configuration framework for customers to explicitly grant additional privileges if needed for their specific use cases, but these extensions require administrative approval and are logged for audit purposes Our system architecture follows defense-in-depth principles with multiple layers of access controls surrounding the LLM.

Example Response 3

No, our current LLM implementation does not limit privileges by default Our solution was designed to maximize flexibility and integration capabilities, allowing the LLM to access various systems and data sources to provide comprehensive responses While this approach enables powerful use cases like retrieving information from connected databases, analyzing documents in storage systems, and interacting with various APIs, we recognize this creates potential security concerns Instead of default restrictions, we provide extensive documentation and configuration options for customers to implement their own privilege limitations based on their security requirements We're currently working on a major update (scheduled for Q3 this year) that will reverse this approach by implementing least-privilege defaults while maintaining the option for expanded access when explicitly configured.

Context

Tab: AI
Category: AI Large Language Model (LLM)