AI Security Startup SherLOCK Provides Technical Support for AISI's 'AI Safety Evaluation Perspective Guide' in Healthcare

SherLOCK, Inc. (Headquarters: Minato-ku, Tokyo; Representative Director and CEO: Teresa Tsukiji; hereinafter 'SherLOCK') has provided technical support for the formulation of the 'AI Safety Evaluation Perspective Guide' in the healthcare domain as a member of the Healthcare Sub-Working Group (SWG) within the Business Demonstration Working Group of the AI Safety Institute (AISI).

Leveraging its specialized knowledge as an AI security startup, SherLOCK contributed to the systematization of specific evaluation methods for practitioners, particularly from the perspectives of security and data quality.

### Background and Objectives While generative AI is expected to bring significant innovation to the healthcare field, it also faces numerous specific challenges such as hallucinations (generation of misinformation), advanced privacy protection, and ensuring security. As discussed at the 'Hiroshima Global Forum for Trustworthy AI' in January 2026, realizing 'Trustworthy AI' is an international priority. SherLOCK participated in this SWG to provide technical insights, aiming to create an environment where companies can guarantee safety from the development stage and balance business value with safety and peace of mind.

### Main Support Provided by SherLOCK In formulating this guide, SherLOCK contributed to the systematization of specific AI safety evaluation methods that practitioners can use immediately, focusing on the following:

1. Provision of Expertise on Evaluation Items and Points from Security and Data Quality Perspectives via AI Red Teaming Tests: - Presented evaluation items and points for tests assuming risks such as intentionally inducing incorrect information through malicious prompts or triggering information leaks. - Provided technical knowledge for building evaluation processes to prevent serious incidents in medical settings. - Defined risks where the quality of data used in RAG (Retrieval-Augmented Generation) undermines the overall product reliability, clarifying evaluation points to ensure the accuracy of medical and health information.

2. Advice on Objective AI Safety Evaluation through Third-Party Assessment: - Provided practical advice and case studies on how AI Red Teaming tests by third parties with specialized knowledge contribute to ensuring corporate reliability in AI safety and security.

### Features of the Guide The guide is designed to be easily utilized even by companies with few specialists, featuring the following: - 5-Phase Evaluation Along the AI Lifecycle: Clearly identifies evaluation points for each phase: Product Design, Model Selection, Implementation, Verification, and Deployment/Operation. - 10-Point Multi-faceted Safety Evaluation Perspectives: Covers everything from universal AI safety perspectives to healthcare-specific risks. Explicitly states specific risks (e.g., direct harm to patient life or health) if evaluation is neglected. - Risk Management Support Linked to Practice: Organizes specific risk assumptions and countermeasures for immediate use by on-site personnel.

### SWG Members (8 Companies/Organizations) Ubie, Inc. (SWG Leader), Awarefy, Inc., Ajinomoto Co., Inc., SB Intuitions Corp., CMIC Holdings Co., Ltd., SherLOCK, Inc., Japan Digital Health Alliance (JaDHA), MICIN, Inc.

### Comment from Teresa Tsukiji, CEO of SherLOCK "In the healthcare domain, where social responsibility is extremely heavy, we are proud to have contributed to the formulation of this AI Safety Evaluation Perspective Guide, which serves as a cornerstone for building trust, alongside AISI and our fellow SWG members. AI safety evaluation is not a barrier to innovation but an essential initiative to accelerate safe social implementation. SherLOCK will continue to support the safe social implementation of AI in various domains, including healthcare, through technical support for advanced AI governance and cutting-edge AI security technologies, such as third-party AI Red Teaming tests."