300 Billion Tokens a Month: Inside NATO’s Massive AI-Powered Cybersecurity Partnership With OpenAI

Artificial intelligence is becoming an increasingly important part of national cybersecurity strategy. As cyber threats grow more sophisticated and digital infrastructure becomes more interconnected, governments and security agencies face the challenge of identifying vulnerabilities, investigating suspicious activity, and responding to attacks at a speed that traditional security processes often struggle to match.
OpenAI's Daybreak cybersecurity initiative represents a significant development in this transition. Through its collaboration with NATO's Communications and Information Agency (NCIA), OpenAI is providing advanced cybersecurity models to support authorized defensive operations across the Alliance. According to OpenAI, NATO has used as many as 300 billion raw input and output tokens per month through this work, illustrating the scale at which AI systems are being integrated into institutional cyber defense.
The initiative extends beyond NATO headquarters. OpenAI has also described cooperation with European cybersecurity organizations, including the European Union Agency for Cybersecurity (ENISA), France's National Cybersecurity Agency (ANSSI), and national cyber defenders in Germany and Poland. Ukraine and Poland have also used OpenAI's models in efforts to identify vulnerabilities and strengthen civilian and network infrastructure.
These developments point toward a broader transformation in cybersecurity. AI is moving beyond its traditional role as an analytical assistant and becoming part of the operational infrastructure used to discover software weaknesses, assess potential attacks, and accelerate defensive engineering. However, its effectiveness will depend on technical reliability, human oversight, responsible deployment, and the ability of governments to maintain control over critical security decisions.
OpenAI Daybreak: Bringing Advanced AI Into National Cyber Defense
Daybreak is designed to give authorized cybersecurity professionals access to advanced AI capabilities for defensive security work. Its applications include reviewing software, investigating unusual system behavior, identifying potential vulnerabilities, validating security findings, and developing and testing corrective measures.
These activities address different stages of the cybersecurity lifecycle.
Traditional vulnerability management often involves several separate processes. Security researchers inspect source code, identify suspicious functions, reproduce potential weaknesses, assess their severity, develop patches, and verify that those patches resolve the underlying problem. Each stage requires technical expertise and can consume considerable time.
AI models can help accelerate parts of this process by examining large codebases, recognizing potentially dangerous programming patterns, generating hypotheses about vulnerabilities, and assisting researchers with testing procedures.
For example, a security team investigating an application may use an AI model to examine how the software handles user input, authentication, memory allocation, or external requests. The model can highlight areas requiring further investigation and help researchers develop tests to determine whether a suspected weakness is exploitable.
This does not make the AI system an independent authority on whether a vulnerability exists. Human researchers must validate the results, reproduce relevant behavior, and determine whether the issue presents a genuine security risk.
The distinction is particularly important in government environments, where an incorrect assessment could disrupt essential services or divert resources away from more serious threats.
Daybreak Blue and Daybreak Red
OpenAI expanded Daybreak in August 2026 by introducing two access categories designed around different cybersecurity activities.
Capability | Primary purpose | Typical defensive applications |
Daybreak Blue | Defensive security operations | Vulnerability discovery, security analysis, incident response |
Daybreak Red | Exploit validation and security testing | Testing weaknesses, validating exploitability, assessing security controls |
The distinction reflects the different requirements of modern security teams.
Defenders need tools that can identify weaknesses before malicious actors exploit them. They also need controlled testing capabilities to understand how vulnerabilities behave under realistic conditions.
Exploit validation is particularly valuable because identifying a suspicious code pattern does not automatically establish that an attacker can exploit it. Testing helps establish whether a weakness can produce unauthorized access, information disclosure, service disruption, or another security consequence.
The responsible use of these capabilities requires authorization, carefully defined testing environments, and appropriate restrictions on potentially harmful operations.
Why NATO Is Investing in AI-Powered Cybersecurity
NATO's digital infrastructure supports communications, coordination, logistics, intelligence exchange, and other functions essential to the Alliance's operations.
Its cybersecurity environment is unusually complex because it involves multiple member states, different national systems, multinational institutions, and networks that must operate together. Protecting this environment requires continuous monitoring and cooperation among specialists with different technical responsibilities.
The collaboration between OpenAI and NATO's Communications and Information Agency is intended to strengthen the protection of Allied networks through access to advanced cybersecurity models.
OpenAI reported that NATO has processed up to 300 billion raw input and output tokens per month as part of the work. Tokens represent the units of information processed and generated by AI models. The figure demonstrates substantial usage, although token consumption alone does not establish the number of vulnerabilities discovered, attacks prevented, or measurable improvements in security.
The scale nevertheless highlights an important operational consideration: deploying AI across a large institution requires more than providing access to a model. Organizations must understand usage patterns, infrastructure requirements, operational costs, data handling, and the reliability of AI-generated findings.
Dr Dylan Browne, General Manager of NCIA, described AI-powered cybersecurity models as a major development for vulnerability discovery and threat detection. He also emphasized the importance of understanding token consumption and model economics as adoption expands, enabling NATO to evaluate competing technologies and make informed procurement decisions.
This is significant because cybersecurity technology is not evaluated solely on technical capability. Governments must also consider long-term operating costs, procurement flexibility, integration requirements, and the strategic consequences of depending on particular technology providers.
Crossed Swords 2026: Testing AI Agents in Cyber Operations
Another important component of OpenAI's work with NATO is its support for the 2026 Crossed Swords exercise, organized by the NATO Cooperative Cyber Defence Centre of Excellence (CCDCOE).
The exercise introduced a training stream in which participants built, deployed, and evaluated AI agents for cyber operations using OpenAI Daybreak within the technical environment.
Unlike a conventional chatbot, an AI agent can be designed to carry out a sequence of related tasks. It may examine information, select appropriate tools, process results, and proceed through a defined workflow toward a particular objective.
In cybersecurity, this could involve coordinating multiple activities, such as reviewing system information, examining suspicious behavior, identifying potential weaknesses, and preparing findings for human evaluation.
The exercise placed participants in operational roles alongside AI agents, with human oversight remaining central to the process.
This approach is important because autonomous capabilities introduce new questions about control and accountability. A system that can execute multiple actions must operate within clearly defined permissions. Human operators need visibility into what the agent is doing, why it is taking a particular action, and what consequences may follow.
A useful operational framework separates AI responsibilities into several stages:
Information collection: The system examines authorized data, logs, code, and security observations.
Analysis: AI identifies patterns and develops potential explanations for suspicious activity.
Recommendation: The system proposes investigative steps or defensive actions.
Validation: Security professionals assess the evidence and reproduce relevant findings.
Remediation: Authorized teams implement and test corrective measures.
This model allows organizations to benefit from automation without treating AI-generated conclusions as unquestionable decisions.
The lessons from exercises such as Crossed Swords can help NATO and allied institutions understand how AI agents behave in complex environments, where operational coordination and human judgment remain essential.
OpenAI's European Cybersecurity Partnerships
OpenAI's Daybreak initiative extends beyond NATO. The company has reported providing cybersecurity model access to European defenders, including ENISA and national cybersecurity organizations in France, Germany, and Poland.
These partnerships demonstrate how AI-assisted vulnerability research can support public institutions responsible for protecting government systems and essential digital services.
ENISA Identifies Four Software Vulnerabilities
One notable example involves the European Union Agency for Cybersecurity, which used OpenAI's models to identify four vulnerabilities in software used across EU institutions.
Fixes were published for all four vulnerabilities:
CVE-2026-73432
CVE-2026-73431
CVE-2026-73405
CVE-2026-73374
The findings illustrate a practical application of AI in software security. Instead of relying exclusively on conventional manual review, researchers can use advanced models to help examine software and identify areas that deserve closer inspection.
A November 2026 joint report from ENISA and CERT-EU provides additional context. It describes a review of 27 software projects using frontier AI models, including OpenAI models accessed through Daybreak and tools such as the Codex security plug-in.
Across the models tested, software maintainers confirmed 95% of the 88 potential issues assessed as genuine.
This result is particularly relevant because it highlights the importance of combining automated analysis with human expertise. AI can help identify possible problems, but software maintainers and security specialists remain responsible for determining whether findings are legitimate and how they should be addressed.
The approach also offers a potential efficiency advantage for organizations maintaining large software portfolios. AI-assisted review can help teams prioritize suspicious code, investigate possible weaknesses, and direct limited engineering resources toward issues requiring attention.
CERT Polska and MikroTik RouterOS
CERT Polska has also used OpenAI models to help identify six vulnerabilities in MikroTik RouterOS, a network operating system used in networking equipment.
The work involved researchers directing the investigation, testing findings on real systems, and verifying their impact before coordinating disclosure with the vendor.
MikroTik subsequently released fixes, which CERT Polska confirmed prevented the attacks it had observed.
This case demonstrates why vulnerability discovery must be connected to responsible disclosure and remediation.
Identifying a weakness is only the beginning. Security researchers must establish its impact, communicate the findings to the affected vendor, allow appropriate time for corrective action, and verify that the resulting updates address the problem.
AI can accelerate the technical investigation, but the broader security process still depends on coordination between researchers, software developers, vendors, and network operators.
Protecting Ukraine's Civilian Infrastructure With AI
Ukraine's cybersecurity requirements have been shaped by persistent cyberattacks affecting civilian infrastructure and essential services.
OpenAI has described the Ministry of Digital Transformation of Ukraine as using Daybreak to strengthen the cyber defense of civilian systems.
The initiative focuses on helping Ukrainian teams identify software vulnerabilities and develop and test fixes more quickly.
Civilian infrastructure is an important target for cyber defense because disruptions can affect services far beyond government technology departments. Telecommunications, energy, healthcare, transportation, financial services, and public administration all depend on interconnected digital systems.
A vulnerability in one component may create exposure across multiple organizations, particularly when software products are widely deployed or integrated into critical operations.
AI-assisted security research can help defenders examine these systems more efficiently, but it must be accompanied by strong safeguards. Sensitive infrastructure requires strict access controls, reliable testing procedures, secure handling of technical information, and human authorization for consequential changes.
The Ukrainian example highlights the potential value of AI in environments where defensive teams must respond to persistent threats while maintaining the availability of essential services.
How AI Is Changing Vulnerability Discovery and Incident Response
The growing use of AI in cybersecurity is connected to a fundamental change in how software security can be approached.
Traditional security tools are effective at detecting known signatures, suspicious activity patterns, and previously identified vulnerabilities. However, new software weaknesses frequently require deeper reasoning about how different components interact.
Advanced AI models can assist with several technical tasks.
Source-code analysis: Models can inspect programming logic and highlight potentially unsafe operations, including weak input validation, insecure authentication, and incorrect permission handling.
Vulnerability research: AI can help researchers develop hypotheses about possible weaknesses and identify code paths that deserve additional investigation.
Exploitability assessment: Within authorized environments, models can assist with designing tests to determine whether a suspected vulnerability has practical security consequences.
Patch development: AI can help engineers understand the affected code and propose possible corrective changes.
Incident investigation: Models can help analysts examine logs, organize technical observations, and develop explanations for unusual system behavior.
Security documentation: AI can assist with producing technical summaries, remediation guidance, and structured reports for development and security teams.
The greatest value may come from connecting these capabilities into a continuous workflow rather than using AI for isolated tasks.
For example, a vulnerability identified during code review could be tracked through validation, severity assessment, patch development, and regression testing. AI could assist at several stages, while authorized professionals retain control over decisions and implementation.
This creates the possibility of reducing the time between discovering a weakness and deploying an effective correction.
However, AI-generated patches must be reviewed and tested carefully. A change that resolves one vulnerability could introduce another defect, weaken compatibility, or affect unrelated functionality.
The Risks of AI-Driven Cyber Defense
Despite its potential, AI does not eliminate the fundamental challenges of cybersecurity.
Advanced models can produce incorrect conclusions, overlook important context, or generate technically plausible but ineffective recommendations. Their performance can also vary depending on code quality, available information, system complexity, and the nature of the task.
Several risks deserve particular attention.
Risk | Potential consequence | Required safeguard |
Incorrect vulnerability findings | Security teams waste resources investigating false positives | Human validation and reproducible testing |
Missed vulnerabilities | Dangerous weaknesses remain undetected | Multiple testing methods and independent review |
Excessive agent permissions | AI systems perform unauthorized or harmful actions | Least-privilege access and approval controls |
Sensitive data exposure | Confidential code or infrastructure information is compromised | Secure deployment and strict data governance |
Unverified AI-generated patches | New defects or operational disruptions | Code review, regression testing, and staged deployment |
Vendor dependence | Critical operations become tied to one provider | Procurement oversight and alternative solutions |
Another concern is the possibility that attackers could use similar AI capabilities to improve their own operations.
Tools that help defenders analyze software may also be misused to investigate vulnerabilities or develop attack techniques. This makes access management, monitoring, usage restrictions, and responsible deployment essential.
The goal should not be to prevent legitimate security research from using advanced AI. Rather, governments and technology providers must establish systems that support authorized defensive work while reducing opportunities for misuse.
The Economics of AI Cybersecurity at Government Scale
The reported consumption of up to 300 billion tokens per month by NATO raises an important question about the economics of large-scale AI deployment.
Token-based pricing can make model usage measurable, but it does not fully describe the cost of operating an AI cybersecurity program.
Organizations must also consider:
Model access and infrastructure expenses.
Integration with existing security systems.
Training and technical support for personnel.
Data protection and regulatory requirements.
Human validation and remediation costs.
Continuous evaluation of model performance.
Long-term procurement and supplier arrangements.
For public institutions, these considerations are especially important because cybersecurity budgets must support predictable and sustainable operations.
The comments from NCIA General Manager Dr Dylan Browne indicate that NATO intends to study consumption patterns and the underlying economics of AI models as adoption grows.
That understanding can help institutions determine which workloads benefit from advanced models, where less expensive systems may be sufficient, and how different providers compare in performance and cost.
The objective should be measurable security improvement rather than maximizing AI usage.
The Future of NATO's AI Cyber Defense Strategy
The collaboration between OpenAI and NATO reflects a wider shift toward AI-supported security operations across governments and critical infrastructure organizations.
Over time, AI systems may become more deeply integrated into vulnerability management, software development, threat intelligence, incident response, and security testing.
Several developments are likely to influence this transition.
More coordinated AI agents: Specialized agents may work together on defined defensive tasks, allowing security teams to coordinate research, analysis, and documentation more efficiently.
Continuous software security: AI-assisted code review could become a more regular part of software development, helping organizations identify weaknesses before applications are deployed.
Improved human-machine collaboration: Training exercises will help establish how operators supervise AI systems, verify their recommendations, and maintain accountability.
Stronger procurement standards: Governments may increasingly evaluate AI models based on security performance, reliability, cost, data governance, and compatibility with existing systems.
Greater international cooperation: Shared research and coordinated vulnerability disclosure can help allied institutions improve security across interconnected networks.
The challenge will be ensuring that technological progress strengthens institutional resilience without creating new dependencies or reducing human control over critical operations.
AI Is Becoming a Strategic Asset for NATO Cybersecurity
OpenAI's Daybreak collaboration with NATO demonstrates how advanced AI is moving into the operational foundations of national and multinational cybersecurity.
The reported use of up to 300 billion monthly tokens, participation in the Crossed Swords 2026 exercise, and partnerships involving ENISA, CERT Polska, and Ukraine illustrate the growing interest in AI-assisted vulnerability discovery and defensive security engineering.
The examples also reveal an important principle: effective AI cybersecurity depends on combining machine capabilities with human expertise. Models can accelerate analysis, help identify software weaknesses, and support remediation, but security professionals must validate findings, authorize consequential actions, and remain accountable for outcomes.
For NATO and allied governments, the long-term objective is not simply to adopt more powerful AI models. It is to build secure, measurable, and sustainable systems that improve the ability to protect essential digital infrastructure.
As Dr. Shahid Masood and the expert team at 1950.ai examine the intersection of artificial intelligence, cybersecurity, and emerging technologies, the broader significance is clear: AI is becoming an important component of national resilience, economic stability, and international security.
Key Takeaways
OpenAI is providing NATO's core institutions with access to advanced cybersecurity models through Daybreak and its collaboration with NCIA.
NATO has used up to 300 billion raw input and output tokens per month, according to OpenAI.
Daybreak supports authorized security activities, including code review, vulnerability discovery, exploit validation, and incident response.
NATO's Crossed Swords 2026 exercise included training on building and evaluating AI agents for cyber operations, with human oversight emphasized.
ENISA used OpenAI models to identify four software vulnerabilities, with fixes published for all four.
A joint ENISA and CERT-EU report described a review of 27 software projects, with maintainers confirming 95% of 88 assessed potential issues as genuine across the models tested.
CERT Polska used OpenAI models to help identify six MikroTik RouterOS vulnerabilities and coordinated their disclosure with the vendor.
Ukraine's Ministry of Digital Transformation is using Daybreak to support the protection of civilian infrastructure.
Human validation, secure access, responsible disclosure, and procurement oversight remain essential to effective AI-powered cyber defense.
Future Outlook
The next phase of cybersecurity will increasingly involve collaboration between human specialists and AI systems capable of processing complex technical information at scale. NATO's experience with Daybreak may provide useful lessons for governments seeking to integrate advanced models into sensitive security environments.
The defining measure of success will be whether these systems help organizations identify genuine threats earlier, resolve vulnerabilities faster, and maintain stronger protection of critical infrastructure. AI can provide defenders with new capabilities, but trust, accountability, and operational discipline will determine their lasting value.
Further Reading / External References
Strengthening NATO and allied cyber defenses with AI
OpenAI is giving NATO access to its Daybreak cybersecurity AI models





Comments