OpenAI and Poland's NASK Sign AI Safety Research Agreement

Poland's NASK and OpenAI sign a research agreement

OpenAI's October 23 announcement says it has signed an agreement with NASK National Research Institute in Poland to explore AI safety, security and technical research. NASK's AI Safety Research Center serves as Poland's AI Safety Institute. The stated work includes evaluating more capable systems, finding risks and improving safeguards. An agreement to explore cooperation is a starting point: the source does not announce a completed evaluation protocol, a deployed national standard or a new law. This October 26 issue records the announcement after its original date.

The partnership matters because safety claims made by a model developer should be tested against the needs of institutions that protect a particular language, public services and information environment. NASK brings those local responsibilities; OpenAI brings access to systems and technical staff. The value of the arrangement will depend on whether researchers can define shared tests, disclose what was measured and act on weaknesses. The announcement names broad research areas but does not publish a schedule, budget, access rules or independently scored results.

What the earlier cyber work does and does not show

The OpenAI announcement points to a separate July collaboration with CERT Polska, which operates within NASK. CERT Polska received access to cyber-capable models through OpenAI's Daybreak program. OpenAI says the team used its models to help discover six vulnerabilities in MikroTik RouterOS, including flaws that could be chained to take control of affected devices without authentication. It says the vendor issued fixes and CERT Polska confirmed that those fixes prevented the observed attacks. The number six belongs to the earlier vulnerability work, not to the new agreement's future results.

That example is evidence that a defender can use the systems in a controlled investigation. It does not establish that the same model will reliably find every vulnerability or that all Polish public services are now protected. A vulnerability count describes findings in one product and period. A claim about national resilience needs a wider denominator: what systems were tested, how issues were reported, whether patches reached users and whether regressions were checked. None of those broader outcomes is quantified in this announcement.

Questions a joint safety program should answer

A useful evaluation program would state the model version, allowed tools, language of prompts and failure definition before tests begin. Polish language and public-sector workflows may expose errors that an English benchmark does not. Researchers should also distinguish harmful capability tests from ordinary deployment monitoring. Those are editorial criteria for judging the partnership, not claims that NASK and OpenAI have already published such a protocol.

The partners could publish methods that other safety institutes can reproduce while keeping exploit details controlled when disclosure would harm defenders. They should distinguish model behavior from network, identity and human approval controls. If a test fails, a reader needs to know whether the model was changed, an external safeguard was added or the task was narrowed. Success should be measured by repeatable protection, not by the prestige of either signatory.

This is therefore a concrete collaboration announcement with one referenced cyber example and substantial unanswered operational questions. It is not a certification of OpenAI systems, a guarantee against synthetic content or a completed international standard. The primary record is OpenAI's October 23 statement; this article's publication date is October 26.