Summary

• NASK and OpenAI will collaborate on research into advanced AI risks, technical testing and safeguards.
• Earlier work through the Daybreak programme helped CERT Polska discover six MikroTik RouterOS vulnerabilities, which the manufacturer patched.
• The agreement does not disclose a research budget, evaluation timetable or independently verified results from the new programme.

Poland’s NASK National Research Institute has signed an agreement with OpenAI to expand research into artificial intelligence safety and cybersecurity, linking public-sector technical expertise with the development and evaluation of advanced AI systems.

The agreement, disclosed on 23 October, provides a framework for examining emerging risks, testing safeguards and investigating how increasingly capable models can contribute to the protection of digital infrastructure.

Researchers from NASK’s AI Safety Research Center will participate in the collaboration. The centre also fulfils the role of Poland’s AI Safety Institute, contributing national research expertise to wider efforts to understand the behaviour of advanced models.

The partnership follows earlier practical cooperation between OpenAI and CERT Polska, the national computer emergency response team operating within NASK. Using access provided through OpenAI’s Daybreak programme, Polish researchers helped identify six vulnerabilities in MikroTik RouterOS, demonstrating one potential application of advanced models to defensive security research.

Neither organisation has disclosed the budget, duration or detailed publication schedule for the new research agreement. The announcement establishes cooperation rather than a completed model evaluation, independent certification scheme or formal regulatory approval process.

Cybersecurity provides an early example

CERT Polska received access to advanced OpenAI models for cybersecurity research in July 2026. The team subsequently used that capability while investigating MikroTik RouterOS, software used to operate networking equipment.

Researchers identified six vulnerabilities, including weaknesses that could be combined to compromise certain devices without authentication. MikroTik supplied fixes, and the Polish team subsequently confirmed that the patches blocked the attack methods identified during its work.

The investigation provides a concrete example of AI assisting technical researchers with a problem requiring detailed software analysis. It does not establish that a model independently discovered and verified every weakness without human intervention.

Vulnerability research involves several stages, including examining software behaviour, identifying suspicious conditions, reproducing a potential flaw and determining whether it can be exploited reliably.

When researchers use AI during that process, they still need to establish which findings are genuine and whether proposed explanations correspond to actual system behaviour.

Incorrect results can consume specialist time or distract teams from more serious weaknesses, while a model may overlook vulnerabilities that do not resemble the examples it has encountered previously.

Responsible disclosure also requires cooperation with the software developer, allowing affected products to be corrected before technical information creates unnecessary additional exposure.

The MikroTik investigation therefore illustrates both the usefulness of advanced AI and the continuing importance of expert verification, coordinated remediation and appropriate handling of sensitive findings.

Advanced capabilities create new security questions

AI systems are increasingly able to write code, interact with software tools, interpret technical documents and complete tasks involving several connected steps.

Those abilities may improve defensive cybersecurity operations, but they also create opportunities for misuse. A model capable of assisting authorised vulnerability research could potentially help an attacker examine similar weaknesses, depending on the access permitted and safeguards applied.

Research into AI safety must consequently examine what systems can do, the circumstances in which harmful behaviour becomes possible and whether controls remain effective when models operate through external tools.

Testing a chatbot in isolation is not equivalent to evaluating an agent connected to databases, software development environments or operational systems. The information available to the model and the permissions granted to it can substantially alter the consequences of an error.

AI agents also encounter information that may originate outside a trusted environment. Malicious instructions embedded in documents, websites or other retrieved material can attempt to change how an assistant behaves.

Evaluations therefore need to examine complete workflows, including whether a system distinguishes authorised user instructions from untrusted content encountered during an investigation.

NASK and OpenAI have described the development of stronger safeguards and more effective methods for identifying threats among the objectives of their cooperation. The detailed testing programmes have not yet been published.

Public research broadens the evaluation process

Frontier-model developers control much of the infrastructure and technical information associated with their products. Independent institutions may need specialist access to investigate capabilities that are not readily observable through ordinary commercial interfaces.

Collaboration can make such research possible, although the value of a partnership depends on the conditions under which tests are performed and results are shared.

NASK brings experience in cybersecurity, public digital services and threats affecting Poland’s information environment. Its researchers have also investigated harmful uses of synthetic content and the reliability of AI systems operating in particular linguistic contexts.

Language presents a practical testing consideration because models do not necessarily behave consistently across all languages. Safeguards developed and evaluated predominantly through English-language examples may produce different results when a user interacts in Polish or another language.

Public researchers can contribute local knowledge about the services, risks and communication patterns relevant to particular countries.

However, participation by a national institution does not automatically establish independent oversight. Research access agreements, confidentiality restrictions and publication rights can influence the extent to which outside experts are able to examine findings.

The organisations have not published enough information about those arrangements to determine precisely how the new programme will balance commercial confidentiality with public-interest research.

OpenAI has existing relationships with the UK’s AI Security Institute and the US Center for Advancing Standards and Innovation for Super Intelligence. The NASK agreement extends those international collaborations, adding Polish expertise to work concerned with AI capabilities and risks.

Shared evaluation methods remain a goal

The organisations have also identified international technical standards as an area where cooperation could become useful.

Comparable methods for evaluating advanced models would allow researchers to distinguish genuine differences in system behaviour from variations caused by separate testing procedures.

Such methods could address capability assessments, misuse risks, the robustness of safeguards and the recording or reporting of significant security incidents.

Developing reliable comparisons will be demanding because the outcome of an AI evaluation can depend on the model version, the instructions provided, the tools available and the criteria used to judge a successful result.

A system may perform well under controlled conditions while behaving differently when connected to live operational software. An evaluation should therefore identify the environment in which a result was obtained and the limits of what can be concluded from it.

Testing can support procurement and governance decisions, but it cannot demonstrate that a model will be safe in every possible application. Organisations deploying AI still need controls appropriate to the systems and information with which it interacts.

The newly signed agreement gives NASK and OpenAI a basis for further research, but its significance will ultimately depend on the technical work conducted and the results made available to others.

The earlier CERT Polska investigation shows that the relationship has already produced identifiable cybersecurity findings. Whether the broader safety programme delivers similarly verifiable results will become clearer as its testing methods, research outputs and practical safeguards are disclosed.