On October 23, 2026, OpenAI announced an agreement with Poland’s NASK National Research Institute to explore collaboration on AI safety, security and technical research.
The announcement comes at a time when AI developers, cybersecurity researchers and policymakers are paying closer attention to systems that behave unexpectedly, exploit weaknesses in software or fail to remain within their intended operating limits.
The partnership will bring together OpenAI’s technical expertise and NASK AI safety research capabilities to study increasingly capable AI systems, identify potential vulnerabilities and strengthen safeguards.
NASK’s AI Safety Research Center serves as Poland’s Artificial Intelligence Safety Institute.
What exactly have OpenAI and NASK agreed to do?
OpenAI and NASK have signed an agreement to explore collaboration across three connected areas: AI safety, security and technical research.
The organisations intend to examine how increasingly capable AI systems can be evaluated, how potential vulnerabilities can be discovered and how safeguards can be strengthened.
NASK brings experience in research related to AI safety, detection of AI-generated content, malicious uses of AI and the challenges of applying safeguards across multiple languages.
Its work also connects technical research with the protection of public services and Poland’s digital information environment.
OpenAI, meanwhile, develops advanced AI systems and conducts research into their capabilities, limitations and potential risks.
Combining these areas of expertise could help researchers examine AI systems from different perspectives.
A model developer may understand how a system was trained and how its safeguards were designed, while an independent research institution can contribute additional testing expertise and challenge assumptions about how those safeguards perform.
This is more than a technology partnership. It reflects a broader shift in the AI industry: as companies build systems that can act, rather than merely respond, testing those systems’ behaviour and security is becoming just as important as improving their capabilities.
Why AI agents are raising new safety concerns?
To understand why this partnership matters, we need to understand how AI agents differ from conventional chatbots.
A conventional chatbot typically responds to a user’s prompt. An AI agent can go further by using tools, interacting with software, executing code, gathering information and completing multiple steps towards a goal.
For example, an AI agent instructed to research a company might search websites, organise findings, compare information and prepare a report.
A more capable agent with authorised access to business systems might also update records, send messages or trigger workflows.
This ability to act makes AI agents useful. It also creates additional risks. An agent may misunderstand an instruction, choose an inappropriate tool, access information it should not use, execute an unsafe action or fail to recognise that it has reached a boundary it must not cross.
The problem becomes more serious when an agent has access to sensitive information, external networks, administrative permissions or other tools that can affect real systems.
What does it mean when an AI agent behaves unexpectedly?
It does not necessarily mean the AI has developed intentions or consciousness.
Unexpected behaviour can result from several technical problems such as
1. Misinterpreting objectives: The system pursues a goal in a way that does not match what its operator intended.
2. Weak access controls: The agent can access tools or resources beyond what it needs for its task.
3. Prompt injection: Malicious instructions embedded in a webpage, document or other input attempt to manipulate the agent.
4. Inadequate monitoring: Security systems fail to detect or record some actions performed by the agent.
5. Insufficient testing: The system encounters a situation that was not adequately covered during evaluation.
These are different problems and require different solutions. Better instructions alone cannot compensate for excessive permissions, and stronger access controls cannot guarantee that every model decision will be correct.
The important shift is that AI safety now has to address not only what a model says, but also what it does, which systems it can access and how its actions are monitored.
OpenAI’s partnership with Poland’s NASK reflects a growing recognition that AI safety requires collaboration between technology developers, independent researchers, cybersecurity professionals and public institutions.
The partnership builds on an earlier cybersecurity collaboration The new agreement is not the first connection between OpenAI and Polish cybersecurity researchers.
According to OpenAI, CERT Polska, the national computer emergency response team operating within NASK, gained access to OpenAI’s cyber-capable models through its Daybreak initiative in July 2026.
The team used the models to help identify six vulnerabilities in MikroTik RouterOS, software used in networking equipment. Some of the vulnerabilities could be combined to take control of affected devices without authentication.
OpenAI reported that the vendor released fixes and that CERT Polska confirmed the fixes prevented the attacks it had observed. This example illustrates a more constructive side of advanced AI security capabilities.
AI can potentially help defenders inspect software, identify suspicious patterns, investigate weaknesses and accelerate vulnerability research.
The same general capabilities, if misused, can also help attackers find weaknesses or develop harmful techniques. The difference often lies in the context, permissions, safeguards and objectives governing how the technology is used.
That creates a challenge for AI developers: they must improve the ability of systems to assist legitimate security work while reducing the risk that those capabilities will be misused.
Collaboration with cybersecurity institutions can help researchers investigate that balance. It also demonstrates why AI safety and cybersecurity, while closely connected, are not identical.
Cybersecurity focuses on protecting systems, networks and information from attacks. AI safety addresses a broader set of questions about how AI systems behave, follow instructions, handle uncertainty and remain within acceptable boundaries. Their overlap is growing as AI systems gain the ability to use digital tools.
What should we watch next?
The next stage of the OpenAI–NASK partnership will be more informative than the announcement itself. The most important developments to monitor are:
1. Research priorities: Which AI safety questions will the two organisations investigate first?
2. Evaluation methods: Will they develop or apply specific methods for testing advanced AI systems and agent behaviour?
3. Independent access: What access will researchers receive to the systems they evaluate, and how will their independence be protected?
4. Public findings: Will the organisations publish research reports, evaluation results or recommendations that other institutions can examine?
5. Measurable improvements: Will the collaboration lead to documented changes in safeguards or better methods for detecting vulnerabilities?
6. International coordination: Will the work contribute to more consistent approaches to AI evaluation, security and incident reporting across countries?
OpenAI’s partnership with Poland’s NASK reflects a growing recognition that AI safety requires collaboration between technology developers, independent researchers, cybersecurity professionals and public institutions.
The partnership builds on an earlier cybersecurity collaboration The new agreement is not the first connection between OpenAI and Polish cybersecurity researchers.
At the time of the announcement, OpenAI had not publicly specified a detailed project timetable, funding arrangement or complete list of deliverables. It is therefore too early to judge the partnership by its eventual technical impact.
