Introducing Hirundo, the World’s First*1 Commercial Machine Unlearning Platform That Removes Generative AI Hallucinations and Information Leakage Risks from Model “Memory”
April 21, 2026
Asgent, Inc.
Asgent, Inc. (Head Office: Chuo-ku, Tokyo; President and Representative Director: Takahiro Sugimoto; TSE Standard: 4288; hereinafter “Asgent”) has entered into a partnership with HIRUNDO AI LTD (Head Office: United Kingdom; CEO and Co-Founder: Ben Luria; hereinafter “Hirundo”), an AI security startup. Beginning in mid-May 2026, Asgent will offer Hirundo's machine unlearning technology in Japan. This technology selectively removes from the “memory” (learned content) of generative AI models information that causes risk.
As the use of generative AI expands into the core of corporate activity, “AI hallucinations,” which generate information not based on facts, and the risk of sensitive information being mixed into models are becoming major management risks that may affect a wide range of corporate activity. This includes zero-tolerance fields where mistakes are unacceptable, such as management decision support, customer response, quality control in manufacturing, and research and development.
In the "Top 10 Information Security Threats 2026 for Organizations" published by IPA,*2 "cyber risks associated with the use of AI" appeared for the first time and ranked No.3. This clearly indicates the seriousness of the issue and the urgent need for companies to establish governance for generative AI.
Against this background, machine unlearning, which selectively removes the influence of specific data and unwanted knowledge from trained AI models, has in recent years attracted attention in both research and development from the perspectives of privacy protection and measures for sensitive information. Since 2023, comprehensive surveys have been published in succession, and in 2025 a systematic overview focused on LLMs also appeared.*3 As a result, machine unlearning is increasing its presence as one of the important technologies in the generative AI era.
Conventional measures such as retraining (full reset), RAG, and guardrails are effective for output control and the supplementation of external knowledge. Hirundo differs from these approaches in that it directly addresses the influence itself remaining inside trained models.
Hirundo provides proprietary technology that acts directly on the weight parameters of trained AI models and pinpoints and removes only the influence of specific concepts. Its key feature is that it can reduce internal model risks quickly and at low cost without retraining. In addition, it contributes to reducing the vulnerability of generative AI to jailbreak attacks such as prompt injection and to improving resistance to bias in outputs. At present, unlearning technologies implemented and provided in commercial environments are still extremely limited, and according to Asgent's research, Hirundo is the world’s first commercial machine unlearning platform.*1
The platform is also available for OEM provision and can be incorporated as an unlearning foundation into domestic LLMs/SLMs and various AI applications, thereby supporting the advancement of Japan’s AI development ecosystem. Going forward, expansion to a comprehensive security foundation for generative AI environments is also planned.
Key Features
・High speed and low cost: Significant reduction in time and cost compared with retraining
・Performance retention: According to Hirundo’s verification, changes in accuracy on major benchmarks are typically less than 1%, including improvements
・Transparency: Visualization of performance indicators before and after modification
・Elimination of root risk: Rather than relying on external control, the platform mathematically disconnects internal model associations and improves jailbreak resistance
Main Areas of Use
・Measures against hallucinations in fields requiring high reliability, such as medical care, legal services, and finance
・Improved safety of AI for management decision support (including AI for boards of directors and management meetings)
・Improved quality and reduced risk in customer response AI
・Use in quality control and zero-tolerance environments in manufacturing
・Unlearning foundation middleware for domestic LLM/SLM and AI application developers (OEM provision)
*1 Based on research by Asgent, as of April 20, 2026. Based on publicly available information, comparison was made among domestic and overseas vendors that commercially provide dedicated machine unlearning platforms for trained AI models that remove the influence of unwanted or sensitive learned signals without assuming retraining.
*2 https://www.ipa.go.jp/security/10threats/10threats2026.html
*3 https://arxiv.org/abs/2305.07512
https://arxiv.org/abs/2306.03558
https://arxiv.org/abs/2503.01854
※文中に記載の会社名、商品名は各社の商標または登録商標です。