GUARD-D-LLM: An LLM-Based Risk Assessment Engine for the Downstream uses of LLMs

TL;DR

This paper introduces GUARD-D-LLM, an innovative LLM-based risk assessment engine that identifies, ranks, and mitigates risks in downstream LLM applications, addressing a critical gap in practical AI risk management.

Contribution

It presents a novel risk assessment framework leveraging multiple intelligent agents to evaluate and mitigate risks specific to downstream LLM uses, filling a gap in existing research.

Findings

Successfully identifies and ranks risks in LLM applications

Provides targeted mitigation suggestions for specific use cases

Facilitates early risk detection and mitigation in LLM development

Abstract

Amidst escalating concerns about the detriments inflicted by AI systems, risk management assumes paramount importance, notably for high-risk applications as demanded by the European Union AI Act. Guidelines provided by ISO and NIST aim to govern AI risk management; however, practical implementations remain scarce in scholarly works. Addressing this void, our research explores risks emanating from downstream uses of large language models (LLMs), synthesizing a taxonomy grounded in earlier research. Building upon this foundation, we introduce a novel LLM-based risk assessment engine (GUARD-D-LLM: Guided Understanding and Assessment for Risk Detection for Downstream use of LLMs) designed to pinpoint and rank threats relevant to specific use cases derived from text-based user inputs. Integrating thirty intelligent agents, this innovative approach identifies bespoke risks, gauges their severity, offers targeted suggestions for mitigation, and facilitates risk-aware development. The paper also documents the limitations of such an approach along with way forward suggestions to augment experts in such risk assessment thereby leveraging GUARD-D-LLM in identifying risks early on and enabling early mitigations. This paper and its associated code serve as a valuable resource for developers seeking to mitigate risks associated with LLM-based applications.