On October 15, 2026, Tokyo-based technology accelerator TECHFUND announced a significant milestone in AI-driven cybersecurity. The company's specialized security copilot, Hi AUDIT Agent, achieved a 90.6% success rate on the CyberGym evaluation platform. This performance surpasses the published Claude Mythos 5 score from Anthropic by 6.8 percentage points, marking a notable shift in the competitive landscape of AI-powered security analysis.

The achievement highlights the growing capability of specialized AI agents to handle complex, real-world software vulnerabilities. Rather than relying solely on general-purpose language models, the integration of domain-specific security expertise allows tools like Hi AUDIT Agent to deliver highly accurate and actionable insights directly to development teams.

The CyberGym Benchmark and the Push for Proof-of-Concept Validation

CyberGym is an advanced evaluation platform developed by a research team at the University of California, Berkeley. Unlike traditional benchmarks that rely on static code analysis or theoretical questions, CyberGym utilizes real-world software vulnerabilities to test the practical capabilities of AI agents.

To succeed on the CyberGym benchmark, AI agents must perform the following tasks:

  • Analyze complex software projects to identify active vulnerabilities.
  • Generate working proof-of-concept tests that successfully reproduce the specific vulnerability.
  • Demonstrate a practical understanding of how the vulnerability behaves in a live environment.

The reported success rate of 90.6% measures the agent's performance on these rigorous benchmark tasks. By exceeding the Claude Mythos 5 score by 6.8 percentage points, Hi AUDIT Agent demonstrates a high level of precision in reproducing and validating security risks, which is critical for developers who need to verify that a vulnerability is genuine before dedicating resources to fix it.

Seamless Integration via Model Context Protocol

A key challenge in modern software development is the friction of switching between different tools. To address this, Hi AUDIT Agent is designed to integrate directly into existing developer workflows. This integration is achieved through the Model Context Protocol (MCP).

By leveraging the Model Context Protocol, developers can connect Hi AUDIT Agent to compatible AI coding tools and editors. This setup enables several key capabilities:

  • Developers can request real-time security analysis without leaving their primary coding environment.
  • Teams can examine security findings in context, viewing the exact lines of code associated with a vulnerability.
  • The agent proposes potential fixes that developers can review and apply directly within their existing workflows.

This workflow-centric approach helps development teams transition smoothly from identifying risks to reviewing and implementing potential fixes, reducing the time-to-remediation for critical software vulnerabilities.

TECHFUND's Broader Ecosystem and Corporate Footprint

Founded in 2014, TECHFUND has established itself as a prominent technology accelerator in Tokyo, supporting both early-stage entrepreneurs and established corporate innovation teams. Over its history, the accelerator has supported more than 500 startup and entrepreneurial teams, as well as intrapreneurs at more than 50 major companies.

TECHFUND's corporate partners include major financial and technology institutions such as:

  • Nomura Holdings
  • SoftBank
  • JCB

The accelerator operates through core business units:

  • Hi STARTUP: This division combines lean startup methodologies with agile development practices to help build and scale new ventures.
  • Hi AUDIT: This division provides comprehensive security audits across emerging technology fields, including AI, Web3, and IoT.

The development of Hi AUDIT Agent represents an extension of TECHFUND's audit expertise into the AI copilot space, combining specialized security knowledge with advanced machine learning models to automate and scale security reviews.

Industry Implications for AI-Driven Security

The benchmark results achieved by Hi AUDIT Agent point to a broader trend in the software industry: the shift from general-purpose AI models to highly specialized, domain-specific agents. While general-purpose models show impressive capabilities across a wide range of tasks, specialized agents trained on specific datasets—such as security audits and vulnerability databases—can outperform them on highly technical benchmarks.

By focusing specifically on security analysis, Hi AUDIT Agent bridges the gap between automated vulnerability scanning and manual security auditing. The ability to generate working proof-of-concept tests is particularly valuable, as it reduces false positives and provides developers with clear evidence of a security flaw. As organizations face increasingly complex security threats across AI, Web3, and IoT environments, the adoption of specialized AI copilots is expected to become a standard practice in secure software development lifecycles.