[Paper Review] Current state of LLM Risks and AI Guardrails

Suriya Ganesh Ayyamperumal, Ge, Limin|arXiv (Cornell University)|Jun 16, 2024

Risk and Safety Analysis14 citations

TL;DR

The paper surveys risks in large language models (bias, safety, privacy, hallucinations, non-reproducibility) and analyzes current guardrails and model-alignment approaches, proposing a layered protection framework and the role of open-source tools.

ABSTRACT

Large language models (LLMs) have become increasingly sophisticated, leading to widespread deployment in sensitive applications where safety and reliability are paramount. However, LLMs have inherent risks accompanying them, including bias, potential for unsafe actions, dataset poisoning, lack of explainability, hallucinations, and non-reproducibility. These risks necessitate the development of "guardrails" to align LLMs with desired behaviors and mitigate potential harm. This work explores the risks associated with deploying LLMs and evaluates current approaches to implementing guardrails and model alignment techniques. We examine intrinsic and extrinsic bias evaluation methods and discuss the importance of fairness metrics for responsible AI development. The safety and reliability of agentic LLMs (those capable of real-world actions) are explored, emphasizing the need for testability, fail-safes, and situational awareness. Technical strategies for securing LLMs are presented, including a layered protection model operating at external, secondary, and internal levels. System prompts, Retrieval-Augmented Generation (RAG) architectures, and techniques to minimize bias and protect privacy are highlighted. Effective guardrail design requires a deep understanding of the LLM's intended use case, relevant regulations, and ethical considerations. Striking a balance between competing requirements, such as accuracy and privacy, remains an ongoing challenge. This work underscores the importance of continuous research and development to ensure the safe and responsible use of LLMs in real-world applications.

Motivation & Objective

• Enumerate risk exposures in deploying large language models.
• Evaluate current technical and implementation challenges of guardrails and model alignment.
• Discuss evaluation methods for bias, fairness, safety, and explainability.
• Propose a layered protection model to secure LLM deployments across external, secondary, and internal levels.
• Highlight the role of system prompts, RAG architectures, and openness in tool support for guardrails.

Proposed method

• Review intrinsic and extrinsic bias evaluation methods.
• Discuss agentic LLM safety with testability and fail-safes.
• Present a layered protection model with GateKeeper, Knowledge Anchor, and Parametric layers.
• Describe guardrails via system prompts, RAG, and bias mitigation techniques.
• Summarize open-source guardrail tools and their approaches.

Experimental results

Research questions

• RQ1 What are the primary risks associated with deploying large language models?
• RQ2 What are the current guardrail and model-alignment approaches, and how effective are they across different layers of protection?
• RQ3 How can evaluation metrics for bias, fairness, safety, and reliability be structured for LLM guardrails?
• RQ4 What challenges remain in balancing flexibility, safety, and cost in guardrail design?

Key findings

• LLMs exhibit biases, safety risks, hallucinations, privacy concerns, and non-reproducibility.
• Guardrails are implemented via a layered protection model spanning external, secondary, and internal levels.
• System prompts, retrieval-augmented generation (RAG), and bias mitigation are central guardrail techniques.
• Fairness metrics and responsible AI considerations are crucial in evaluating biases and datasets.
• Open-source tools (Nemo-Guardrails, LlamaGuard, Guardrails AI) offer diverse DSLs and evaluation strategies for guardrails, despite cost and bias concerns.
• Challenges remain in achieving optimal trade-offs between flexibility and safety, testability, and real-world cost.