IN Brief:
- New Zealand’s NCSC has issued guidance on cyber risks from frontier AI-enabled vulnerability discovery.
- The guidance highlights shorter exploitation windows, supply-chain exposure, patch management, and secure use of AI tooling.
- Defence manufacturers face growing pressure to secure embedded software, operational technology, digital design environments, and supplier code.
New Zealand’s National Cyber Security Centre has issued guidance on the cyber risks created by frontier AI-enabled vulnerability discovery, with clear consequences for defence manufacturers, aerospace suppliers, and national-security technology companies built around complex software supply chains.
The guidance addresses a fast-moving technical shift. Frontier AI models are becoming more capable of identifying software vulnerabilities, accelerating work that can help defenders while also reducing the time available to patch exposed systems. Defence suppliers face a particularly awkward version of that problem because many of their products and production environments rely on long-lived software, embedded firmware, specialist tools, and multi-tier supplier code.
Practical controls sit at the centre of the guidance. Organisations using AI for vulnerability discovery are urged to limit access, use development or isolated environments, apply least-privilege service accounts, sandbox activity where appropriate, and understand how AI-as-a-service tools may handle source code, intellectual property, or sensitive system information. The same document reinforces the discipline behind vulnerability management: identify, prioritise, validate, remediate, report, and filter false positives.
For defence industry, the deeper pressure lands on production assurance. A modern military platform is as much a software and data system as a mechanical product. Aircraft, vehicles, ships, sensors, radios, ground stations, satellites, manufacturing execution systems, and design environments all depend on code that may come from internal teams, subcontractors, open-source components, commercial packages, legacy libraries, and bespoke mission applications.
AI-enabled vulnerability discovery changes the tempo around that code. Weaknesses that once required specialist manual effort may become easier to find at scale. Attackers can use AI to triage targets, write exploit variants, interpret technical documentation, and move through exposed systems more quickly. Defenders can use similar tools to improve scanning and remediation, but only where asset visibility, patch governance, and incident response are mature enough to absorb the workload.
Cyber skills are already being pulled into the defence production base, with funding for Winchester’s cyber work showing how training, resilience, and industrial capability now overlap: Winchester cyber funding puts skills into the defence production base. New Zealand’s guidance adds another layer to that same problem. Cyber competence is no longer a support function behind defence manufacturing; it is part of how secure platforms, software-defined systems, autonomous vehicles, and connected factories are built.
Software bills of materials will become more important as exploitation cycles shorten. Defence primes increasingly need visibility into libraries, dependencies, versions, and update routes inside the systems they integrate. That visibility is hard to achieve when suppliers use proprietary stacks, subcontracted code, or embedded devices with limited update support. AI-assisted discovery makes weak inventory discipline more costly because unknown dependencies can quickly become unknown exposures.
Operational technology creates an additional constraint. Defence manufacturers rely on CNC machines, robotics, quality systems, test rigs, industrial control systems, and connected production networks. Many of those environments cannot be patched at the same speed as office IT because downtime can affect output, certification, or safety. Shorter exploitation windows will push manufacturers towards segmentation, monitoring, strict access controls, secure backups, and tested recovery procedures.
The guidance also speaks to AI use inside engineering organisations. Developers and cyber teams will want to use AI to review code, find vulnerabilities, and accelerate security testing. That can improve resilience, but it creates data-handling problems. Source code, model files, vulnerability details, design documentation, and proprietary interfaces may be too sensitive for uncontrolled external tools. Defence suppliers will need governance that allows useful experimentation without leaking information adversaries want.
Smaller suppliers could feel the pressure most sharply. Defence supply chains depend on SMEs with specialist capability in sensors, software, electronics, materials, testing, and integration. Many do not have large cyber teams, yet they may hold export-controlled designs, production data, or access credentials for prime portals. AI-enabled threats could widen the gap between the security expectations placed on those suppliers and the resources available to meet them.
The guidance gives defence industry a practical warning. AI will accelerate both discovery and exploitation, while defenders will be expected to move faster without sacrificing assurance. Manufacturers that treat cyber as a compliance exercise will struggle as military systems become more autonomous, connected, and software-defined. Those that build software assurance, SBOM management, patch discipline, monitoring, and incident recovery into production practice will be better prepared for the next stage of cyber pressure.
