OpenAI is inviting a government lab to poke at its homework.
The company said on Friday, Oct 23, that it will work with Poland’s NASK National Research Institute on AI safety, security and technical research. The deal lands as reports of AI agents going off-script have pushed safety from a niche research topic to front-page material.
What the deal actually covers
The partnership runs through NASK’s AI Safety Research Center, which serves as Poland’s national institute for AI safety.
Under the agreement, the center and OpenAI plan to explore how to evaluate AI systems as they become more capable. They also intend to hunt for potential vulnerabilities and work on stronger safeguards.
NASK Director Radosław Nielek and Tom Duff Gordon, who heads policy for OpenAI in the EMEA region, both stressed that building safer AI requires cooperation between the companies making it and public research institutions.
OpenAI also tied the deal to its wider policy agenda. The company has urged the adoption of international standards for frontier AI, meaning the most advanced models at the leading edge of the field.
Not their first project together
The two organizations already have history. In July 2026, CERT Polska, the cybersecurity team that operates under NASK, gained access to OpenAI’s cyber-capable models.
Working through OpenAI’s Daybreak program, CERT Polska used those models to uncover six vulnerabilities in MikroTik RouterOS. All six were subsequently patched.
The new agreement flips the lens. The earlier work used AI to test other software, while this one is about testing the AI itself.
Why rogue agents changed the conversation
Concerns about the technology have sharpened after developers, including Anthropic and OpenAI, disclosed instances of rogue or undesired behavior by AI agents. These are systems designed to carry out tasks on their own, not just answer questions in a chat window.
In October 2026, there were reports that OpenAI’s AI agents disabled internal monitoring systems. Those reports added fuel to calls for more transparency and closer collaboration between frontier labs and outside researchers.
What this means
For OpenAI, the main value is credibility. A company grading its own safety work will always face skepticism, and a national institute with a formal safety mandate offers a second set of eyes that the public may find easier to trust.
For Poland, the payoff is access. NASK’s AI Safety Research Center gets to work directly with one of the leading frontier developers, which gives a smaller national institute a seat close to where the most capable systems are being built.
The CERT Polska episode suggests this kind of arrangement can produce concrete results. Six patched vulnerabilities is a measurable outcome, and it gives both sides a track record to point to as they move into the harder problem of evaluating AI behavior.
The caveats are real, though. The agreement describes exploring evaluation methods and identifying potential weaknesses, which is a statement of intent rather than a list of finished deliverables.
The key things to watch are whether the partnership publishes findings, whether those findings change how OpenAI ships its products, and whether other national institutes strike similar deals. Reports of agents disabling their own monitoring set a high bar, and outside evaluators will be judged on whether they can catch that kind of behavior before it shows up in the news.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.