NATO and allied agencies use OpenAI models for cyber defense and vulnerability testing

Article Summary

OpenAI is providing NATO and various European cyber agencies access to its Daybreak AI platform to support software vulnerability discovery, threat investigation, and fix development. NATO institutions have utilized up to 300 billion raw input and output tokens monthly, while other partners including Ukraine's Ministry of Digital Transformation, ENISA, and national agencies in France, Germany, and Poland leverage these tools for civilian and institutional network protection. Additionally, OpenAI supported the 2026 Crossed Swords exercise, where participants trained with AI agents for cyber operations while maintaining human oversight protocols.

AI Analysis

## Core Assessment

The transition of generative models from back-office productivity tools to mission-critical infrastructure security is now formalized. OpenAI providing its Daybreak platform to NATO, ENISA, Ukraine's Ministry of Digital Transformation, and national agencies across France, Germany, and Poland demonstrates that institutional resistance to hosted proprietary models in defense contexts is eroding under operational pressure. The reported volume—NATO's Communications and Information Agency consuming up to 300 billion raw tokens monthly—signals deep pipeline integration for automated triage, patch generation, and vulnerability discovery rather than ad-hoc pilot testing. The verified operational utility is real: ENISA confirmed four published CVEs with maintainers validating 95 percent of 88 identified issues, and CERT Polska leveraged the models to find six MikroTik RouterOS flaws.

However, institutional dependence on commercial AI inference for sovereign network defense introduces serious procurement and operational constraints. As NATO’s leadership openly noted, scaling these workflows requires untangling unpredictable token economics and usage volatility. Feeding sovereign telemetry, proprietary source code, and zero-day patch workflows into commercial model interfaces demands strict compliance isolation, continuous verification against model hallucinations, and defense-in-depth protocols. While exercises like Crossed Swords 2026 deliberately enforce human-in-the-loop controls for agentic cyber operations, the defensive advantage will last only until identical automated audit and exploitation capabilities are scaled by adversaries.

## Impact and Advice

* **Enterprise Security Operations Leaders**: Audit static analysis and incident response workflows to identify bottlenecks suitable for assisted code analysis, but mandate strict human verification. While validated vulnerability discovery rates are promising, AI-generated patches must pass full regression testing to prevent introducing secondary defects.
* **Public Sector and Defense IT Procurement Officers**: Establish firm cost-governance guardrails for consumption-based billing models. Monthly utilization reaching billions of tokens will create severe budget variance unless organizations establish workload-specific quotas and clear evaluation benchmarks against competing private-cloud or on-premise solutions.
* **Infrastructure and Product Security Engineers**: Prioritize continuous, automated code remediation on exposed edge and routing equipment. Vulnerability discovery that previously took specialized research teams months is compressing into rapid cycles, drastically shortening the window between flaw discovery and active exploitation.

## Bottom Line

Having an AI assistant catch bugs in your routing firmware is great engineering assistance, but relying on a commercial meter to defend sovereign infrastructure means your security posture is only as stable as your monthly token bill.

Explore this story further