GMO Flatt Security Announces Integration Plan for 'Mythos-class Models' and New 'AI Penetration Testing' Feature for Takumi

GMO Flatt Security Co., Ltd. (President: Yasutaka Ide; hereinafter 'GMO Flatt Security'), a GMO Internet Group company that develops cybersecurity-related businesses for product development organizations with the mission of 'supporting engineers,' announced that it plans to integrate Anthropic's 'Mythos-class models' into its security AI agent 'Takumi byGMO' (hereinafter 'Takumi') as soon as they are released to the public. Additionally, the company announced a new feature, 'AI Penetration Testing,' which maximizes the expected performance in cybersecurity. Through the provision of these features, the company aims to support Japanese companies in adopting frontier AI and improving their cybersecurity. 'Takumi' is a security-specialized AI agent developed by GMO Flatt Security. Since its release in March 2025, it has continuously supported robust implementation in software development organizations as an AI that autonomously performs security diagnostics, including black-box diagnostics (DAST), white-box diagnostics (SAST), and automatic vulnerability correction, by detecting vulnerabilities and creating correction patches. In March 2026, the company released 'Takumi Guard,' which blocks the download of malicious packages in the software supply chain, and 'Takumi Runner,' an execution platform that visualizes the behavior of workflows used in CI/CD environments, supporting software security consistently from development to daily diagnostics. 'Claude Mythos Preview' is a general-purpose AI model announced by Anthropic in April 2026, which claims high performance in the cybersecurity field. In May 2026, Anthropic announced the new model 'Claude Opus 4.8' and stated that it is working to provide 'Mythos-class models' to all customers within a few weeks. While the public release of 'Mythos-class models' can contribute to improving cybersecurity if used correctly, there is also a significant concern that attackers will exploit them. Japanese companies are also required to strengthen their defenses through active AI utilization to counter the significant improvement in the speed and skill of attackers. The integration of 'Mythos-class models' into the various functions provided by 'Takumi' is planned to proceed sequentially after the public release. The target functions include the new 'AI Penetration Testing' feature, security diagnostics and automatic correction functions, and 'Takumi Guard' and 'Takumi Runner' functions. By incorporating the model into these key functions, the company aims to realize a more advanced and comprehensive cybersecurity environment. The new 'AI Penetration Testing' feature goes beyond traditional vulnerability diagnostics, which merely confirm the presence of vulnerabilities, to investigate the actual exploitability more deeply. Furthermore, it scrutinizes the impact of larger attacks caused by combinations of vulnerabilities and verifies whether predetermined goals (such as extracting confidential information from services or service suspension) can actually be achieved.