NTT Develops Sovereign AI for Vulnerability Analysis

NTT Develops Sovereign AI for Vulnerability Analysis
Monday, November 16, 2026
  • •NTT developed AI that analyzes vulnerabilities without sending source code outside the organization’s controlled environment.
  • •The method achieved a 63% success rate on 150 CyberGym tasks, up 15 percentage points from the previous 48%.
  • •NTT aims to help organizations find vulnerabilities, assess their own risks and decide how to address them.

On November 16, 2026, NTT announced a cybersecurity AI technology that analyzes vulnerabilities in a controlled environment without sending the source code under inspection to an external AI service. It combines multiple open-weight models with analysis techniques informed by vulnerability research and NTT’s own workflow controls. In benchmark testing, the success rate rose to 63% from 48% with the previous method, a 15-percentage-point improvement. NTT aims to balance strong analysis capabilities with “sovereignty,” allowing organizations to choose and operate AI themselves.

Using high-performance external AI services can raise concerns about sending confidential source code outside the organization, rising costs and dependence on a particular service. Finding vulnerabilities alone does not reduce risk: organizations must determine which issues affect their systems, prioritize them and address them safely. NTT is pursuing research and development to support the process from discovery through assessment and remediation.

In general AI-based vulnerability analysis, a model reads a program, forms hypotheses about possible vulnerabilities and tests them. AI can repeat work resembling that of specialists autonomously and run many analyses in parallel. Results depend not only on model performance but also on the workflow that guides the work and the tools used for testing. NTT incorporated guidance based on its accumulated research to direct AI toward parts of a program where vulnerabilities may arise. The company says this helps draw out the AI’s capabilities and improve its analysis success rate.

For its evaluation, NTT selected 150 memory-corruption tasks from roughly 1,500 challenges in CyberGym, a benchmark for measuring AI agents’ vulnerability-analysis capabilities. Comparing an existing analysis workflow with the improved method, it recorded a success rate increase from 48% to 63%. NTT also evaluated open-weight models of different sizes and found potential to improve performance by combining medium-sized models. The company plans to integrate the approach with its own analysis technology, which repeatedly varies inputs to a system under inspection and looks for potential vulnerabilities based on abnormal behavior.

NTT plans to develop technology that matches vulnerability information against an organization’s system configuration to identify affected areas and risks that should be addressed first. It also plans to support remediation that takes system operations and business impact into account. Publicly disclosed vulnerabilities do not all affect every organization’s systems, and the need and priority for remediation vary. NTT aims to connect discovery, assessment and remediation to reduce the time and effort required to lower actual risk. The results are scheduled to be exhibited at NTT R&D FORUM 2026.

Subscribe to the AIB Newsletter

The fastest, most reliable AI news, straight to your inbox

Blog posts related to this article