How to choose a Japanese AI security product

For example, a team that wants to reduce SQL injection vulnerabilities in a SaaS product before release needs a different product from a team that wants to organize the large volume of alerts it receives every day. The former primarily needs code analysis or Web security assessments, while the latter needs support for monitoring and investigation. Protecting generative AI itself is another use case, requiring a review of how prompts and external data are handled.

To make comparisons easier, this article covers products and services offered by providers headquartered in Japan. This does not mean that their components or AI models were developed in Japan, or that their data is processed there. The examples below are starting points for selection, not an exhaustive list of all products.

If you want to examine code under development, Hi AUDIT Agent from TECHFUND is an option for requesting code security analysis from a development environment connected via MCP. One possible workflow is to review the flagged locations and proposed fixes, then have developers verify the fixes.

If you want to continuously assess both code and running applications, Takumi byGMO from GMO Flatt Security offers code-based assessments, black-box assessments, and suggested fixes. It also advertises scheduled assessments of GitHub diffs, so teams that prioritize integration with their development workflow may want to look into it.

If you want to assess a Web application from the outside, AeyeScan from Aeye Security Lab is a cloud-based tool that uses AI to automate security assessments. Whether it can crawl pages behind authentication is something to test with your actual settings and target application.

If you cannot send confidential data outside your environment, Purple Agent from CoWorker advertises vulnerability assessments and forensic analysis using an on-premises local LLM. Check the deployment requirements and the operational setup you will need in advance.

For managing employees’ use of generative AI, Layer Base from Layer Security inspects browser inputs and attachments on the device and can warn, replace, or block according to policy. The company also describes controls for supported AI-agent operations. Confirm supported browsers and actions, and assess this separately from runtime protection for AI-enabled applications.

If you want to reduce the burden of security monitoring, CTC-MSS AI SOC from ITOCHU Techno-Solutions supports AI-powered threat detection, investigation, and suggested responses. Although the service is offered by a Japanese company, it uses products from the US company Exaforce as its foundation.

When comparing AI assessment results, it is difficult to judge which is better based solely on the number of findings. An assessment that stopped at the login screen and one that reached pages for different permission levels have examined fundamentally different scopes. For code analysis, also check whether the assessment covered the relevant repositories and languages, as well as the paths from callers through to the database.

Test against the same application and revision, with authorization, and compare whether each product detects known issues, lets you trace the basis for its findings, and can check again after fixes. Deciding who is responsible for verifying findings and making fixes can also help prevent the product from falling out of use after adoption.

Even when a service is offered by a Japanese company, where input data is stored, where inference takes place, and whether external models use the data for training are separate conditions. Before sharing confidential code or logs, check the contract and the latest documentation for retention periods, deletion methods, and subcontractors as well.

Check each data condition separately. AeyeScan states that it stores customer data in AWS's Japan region, while explaining that data on Azure for its generative AI option is automatically deleted after being retained for up to 60 days from use. For our service, we explain that TECHFUND itself does not use the data for training, but the selected external model may use it for training. Storage in Japan, use for training, retention periods, and processing regions are not the same condition.