Daybreak moves into operational defence
OpenAI has detailed how NATO and several European cyber agencies are using its Daybreak system and frontier models for authorised defensive work. NATO’s Communications and Information Agency is providing access across core institutions, with usage reaching as much as 300 billion raw input and output tokens per month. That scale moves AI security agents beyond occasional research assistance into an operational capability supporting software review, investigation, vulnerability validation and the development and testing of fixes.
The most important fact is not the token count by itself. It is the integration of model-assisted analysis into multinational teams responsible for live networks. OpenAI says the 2026 Crossed Swords exercise included a stream in which participants built, deployed and evaluated cyber agents in a technical environment using Daybreak. The exercise kept human operators alongside the agents, an appropriate design where an incorrect exploit, scan or remediation could affect systems outside the intended scope.
Evidence from allied agencies gives the announcement more substance than a hypothetical demonstration. The European Union Agency for Cybersecurity used frontier models in a review of software projects and identified four vulnerabilities for which fixes were published. OpenAI cites a joint ENISA and CERT-EU assessment of 88 potential issues across 27 projects, with maintainers confirming 95 per cent as genuine. Confirmation by project maintainers matters because model-generated security findings often contain duplicates, unreachable paths or overstated impact.
Evidence comes from verified vulnerability work
CERT Polska used OpenAI models to help find six vulnerabilities in MikroTik RouterOS, then coordinated disclosure and verified the vendor’s fixes. Ukraine’s Ministry of Digital Transformation is using Daybreak to support defenders protecting civilian infrastructure. These cases illustrate the strongest role for frontier models today: expanding the search space and accelerating analysis while trained researchers reproduce findings, assess real-world impact and control disclosure. The model supplies leverage, not independent authority to probe or change a target.
At NATO scale, consumption economics become a strategic concern. Three hundred billion tokens a month can represent very different cost and performance outcomes depending on model mix, caching, context design and how many findings survive review. Security leaders need task-level measures such as confirmed vulnerabilities, time to remediation, analyst hours saved and false-positive burden. A raw token volume can prove adoption but cannot establish value, especially when repeated context or poorly scoped agents consume large amounts of compute.
Scale creates procurement and oversight questions
The deployment also raises operational questions about data handling. Vulnerability research may expose source code, configurations, credentials and details of unpatched systems. Agencies need clear tenancy, retention, access logging and residency controls, with separation between organisations collaborating through NATO. Prompts, tool results and exploit artefacts should be classified and protected like other sensitive operational records. Model and tool updates require change control because a behavioural change can affect both detection quality and the risk of unintended actions.
Authorised scope must be enforced outside the model. Network boundaries, allow-lists, rate limits and disposable test environments should prevent an agent from reaching systems merely because a prompt suggests it. Tool permissions need to distinguish passive review from active scanning and exploitation. High-impact actions should require explicit human approval, and every step should generate tamper-resistant evidence linking the operator, model, tools, target and result. Natural-language instructions alone are not a security control.
Human authority remains the boundary
For Australian government and critical-infrastructure teams, the NATO example is relevant even without access to Daybreak. It shows that frontier cyber models can support genuine defensive outcomes when combined with experienced analysts and coordinated disclosure. Procurement should test models on representative code and local operating constraints rather than relying on benchmark claims. Exercises should include compromised agent instructions, false findings, data leakage attempts and the need to stop a run immediately without losing its audit trail.
OpenAI’s announcement marks a maturing phase for AI-assisted cyber defence. The technology is being judged by verified fixes and operational resilience, not only by its ability to describe vulnerabilities. The opportunity is faster discovery and remediation across software too large for human teams to inspect exhaustively. The corresponding obligation is disciplined oversight: narrow scope, reproducible evidence, human confirmation and cost measurement that rewards safer systems rather than the largest volume of automated activity.
A mature programme should also preserve negative results. If an agent investigated a suspected flaw and the maintainer later determined it was unreachable, that outcome belongs in the evaluation record alongside confirmed findings. Otherwise teams will optimise reporting around successes and underestimate review cost. Shared taxonomies for severity, reproducibility and model contribution would help NATO members compare results without disclosing operational secrets. They would also reveal whether improvements come from a stronger model, better tools, more context or increasingly experienced human operators. That evidence is necessary before governments expand autonomous scope or commit to larger recurring inference budgets responsibly.