Menu

Security Products

A Machine Unlearning Platform that reduces risks within generative AI models without retraining
Hirundo

Hirundo is a Machine Unlearning Platform that helps reduce risks such as hallucinations, data leakage, bias, and jailbreaks by addressing unwanted knowledge and undesirable behavior at the model level. Rather than relying solely on external guardrails or output filters, Hirundo makes targeted modifications to the weights and behavior of trained models, helping organizations use generative AI more safely while reducing the time and cost associated with retraining.

Challenges in using generative AI

As the use of generative AI expands across enterprises, hallucinations that produce plausible but incorrect answers, the disclosure of sensitive or personal information derived from training data, biased responses, and attacks such as prompt injection have become major obstacles to production use. In high-trust fields such as management decision support, customer service, manufacturing quality control, research and development, healthcare, legal services, and finance, organizations need mechanisms that continuously maintain output quality and safety.
Conventional guardrails and output filtering are important safeguards, but they do not remove the underlying risks that remain within the model. Retraining and additional fine-tuning can also be effective, but they may require substantial time, cost, and validation effort, making it difficult to respond quickly to issues discovered shortly before release or during production.
Hirundo helps balance the safety and practical use of generative AI by assessing model weaknesses, reducing identified risks through Machine Unlearning, and supporting ongoing protection during operation.

Three key benefits provided by Hirundo

  1. Diagnose: Identify model risks

    Built-in evaluations and red-team testing help identify risks within a model, including hallucinations, personally identifiable information (PII), bias, and prompt injection. Evaluation results before and after modification can be compared, making it easier for both technical and business teams to assess improvements.

  2. Harden: Reduce undesirable behavior within the model

    For open-weight models, Hirundo applies Machine Unlearning to the parameters and behaviors associated with identified risks, reducing unwanted knowledge and vulnerable response patterns. Because it performs targeted modifications rather than retraining the entire model, Hirundo helps improve model behavior while preserving its overall utility.

  3. Protect: Support continuous protection after deployment

    When new risks are identified during operation, organizations can consider improvements without waiting for a lengthy retraining process. For closed-weight and API-based models, Prism adjusts token probabilities at inference time, helping suppress unsafe generation.

Key features

  1. Address risks within the model without retraining

    Fundamentally correcting problems in a trained AI model would ordinarily require retraining or extensive additional tuning. Rather than rebuilding a model from the ground up, Hirundo takes an approach that reduces the influence of unwanted knowledge and undesirable behavior remaining in the existing model. This makes it suitable for improving quality before release and responding quickly to problems discovered in production environments.

  2. Address hallucinations, PII, bias, and vulnerabilities

    Hirundo evaluates common generative AI risks, including hallucinations, the inclusion of PII or confidential information, response bias, prompt injection, and jailbreaks, and addresses the model behavior that causes them. According to information published by Hirundo, testing has demonstrated reductions of up to 85% in successful jailbreaks, reductions of up to approximately 70% in bias, and the removal of fine-tuned PII.

  3. Combine Hirundo with guardrails for multilayered AI security

    Hirundo is not intended to replace guardrails. It is a complementary mechanism that reduces risks within the model that cannot be fully addressed through external controls alone. By combining Hirundo with input and output controls, retrieval-augmented generation (RAG), monitoring, and access management, organizations can enhance generative AI safety through multiple layers of protection.

  4. Evaluate improvements and verify their effectiveness

    Evaluations of the targeted risks and general utility benchmarks are conducted before and after modification. In addition to reducing risks, Hirundo emphasizes minimizing the impact on reasoning capabilities and general knowledge. Published information from Hirundo also presents test results showing that the impact on major benchmarks was kept low.

  5. Supports OEM deployment and integration into existing AI platforms

    In addition to being used as a standalone solution, Hirundo is designed for use as a Machine Unlearning platform integrated into LLMs, SLMs, and AI applications. It is also suitable for organizations seeking to enhance the quality and safety of their proprietary models, domestically developed LLMs and SLMs, and AI services.

Use cases

Improving generative AI safety in high-trust fields

In fields such as healthcare, legal services, and finance, where incorrect answers or information leakage can have serious consequences, both accuracy and safety are essential. Hirundo evaluates and reduces hallucinations, PII, bias, and vulnerability to jailbreaks that remain within the model, supporting the use of generative AI in high-trust operations.


Risk assessment before deployment and improvement before release

If red-team assessments or external testing identify risks before a new AI service is released, Hirundo provides an option for making improvements without relying on retraining. It can support final-stage quality improvements and provide the risk visibility needed to make release decisions.


Rapid response to problems identified in production

If a new vulnerability or undesirable behavior is discovered after a system enters operation, Hirundo can assess the affected areas of the model and support targeted modifications. It is suitable for situations in which continuous improvement is required while avoiding service interruptions or lengthy retraining.


Improving the quality of AI used for management decision support and customer service

AI systems used to support analysis for boards of directors and management meetings, as well as internal and external inquiry services, need to suppress plausible but incorrect or biased answers and maintain consistent response quality. Hirundo supports safer AI operation by assessing model weaknesses, reducing undesirable behavior, and providing protection after deployment.


Integration into domestically developed LLMs, SLMs, and AI applications

In environments with strict quality and safety requirements, it is important to address not only external guardrails but also risks that remain within the model itself. As a Machine Unlearning platform that can be integrated into AI platforms and applications, Hirundo also helps improve the reliability of the domestic AI ecosystem.


Use in zero-tolerance environments such as manufacturing and research and development

In quality control, design support, research and development, and other environments where incorrect answers or information leakage are difficult to tolerate, organizations need a framework for continuously evaluating and improving AI output quality. By reducing risks within the model and providing visibility into evaluation results, Hirundo supports the use of generative AI in operations where errors cannot be tolerated.



Security Products

Go to the list page "Security Products"